Download Cluster Computing: Architectures, Operating Systems, Parallel

Cluster Computing
Architectures, Operating Systems, Parallel Processing &
Programming Languages
Author Name:
Richard S. Morrison
Revision Version 2.3, Tuesday, 2 April 2003
Copyright © Richard S. Morrison 1998 – 2003
This document is distributed under the GNU General Public Licence [39]
Print date: Tuesday, 2 April 2003
Document owner: Richard S. Morrison, [email protected] ✈ +612-9928-6881
Document name: CLUSTER_COMPUTING_THEORY
Stored: (\\RSM\FURTHER_RESEARCH\CLUSTER_COMPUTING)
Revision Version 2.3
大纲
Copyright © 2003
Synopsis & Acknolegdements
发动
My interest in Supercomputing through the use of clusters has been long standing and was
initially sparked by an article in Electronic Design [33] in August 1998 on the Avalon Beowulf
Cluster [24].
Between August 1998 and August 1999 I gathered information from websites and parallel
达到顶点 research groups. This culminated in September 1999 when I organised the collected material
and wove a common thread through the subject matter producing two handbooks for my own
use on cluster computing. Each handbook is of considerable length, which was governed by
the wealth of information and research conducted in this area over the last 5 years. The cover
the handbooks are shown in Figure 1-1 below.
Figure 1-1 – Author Compiled Beowulf Class 1 Handbooks
Through my experimentation using the Linux Operating system and the undertaking of the
University of Technology, Sydney (UTS) undergraduate subject Operating Systems in Autumn
Semester 1999 with Noel Carmody, a systems level focus was developed and is the core
element of this material contained in this document.
This led to my membership to the IEEE and the IEEE Technical Committee on Parallel
Processing, where I am able to gather and contribute information and be kept up to date on the
latest issues.
My initial interest in the topic has developed into a very practical as well as detailed theoretical
knowledge. In Autumn semester 2001 I undertook to tutor the UTS Operating Systems subject
which included guest lectures. This enabled me to further develop my ability to articulate my
knowledge to students with no previous experience of the internals of operating systems or
systems level programming and facilitate their learning.
用。。。连接
Cluster Computing
Page 3
Revision Version 2.3
Copyright © 2003
Focus
This document reviews the current parallel systems theory with respect to Beowulf type
clustering and experimentally investigates certain elements of parallel systems performance.
The investigation work was carried out at the University of Technology, Sydney in Laboratory
1/2122E.
A separate document could be written on developing applications for parallel systems and
currently it would be largely dependant on the target system.
I have however focused on the systems level as it has been apparent over the years that I have
been researching the area that while Scientific Computing drove the HPCC to develop
clustering, Corporations are now requiring the same level of advantages as possible with
Clusters. As such, Operating system vendors are now taking the experience of the HPCC and
integrating these features into their operating systems.
Next generation tools are available now to develop parallel programs and it is envisioned that
parallel systems will be the only model in the future. Before this can happen, standardisation
must be reached (Formal or Pseudo), as this will be an important step to minimise the
transitional cost to parallel software. [Refer to Section 3.4.2 for further detail]
Page 4
Cluster Computing
Revision Version 2.3
Copyright © 2003
Executive Summary
Cluster Computing:
Architectures, Operating Systems, Parallel Processing & Programming Languages
The use of computers within our society has developed from the very first usage in
1945 when the modern computer era began, until about 1985 when computers were
large and expensive.
Two modern age technologies, the development of high-speed networking and the
personal computer have allowed us to break-down these price barriers and construct
cost effective clusters of PCs which provide comparable performance to supercomputers at a fraction of the cost. As PC’s and networks are in common use, this
allows most commercial organizations, governments, and educational institutions
access to high performance super-computers.
The major difference between a network of PC’s and a super-computer is the software
which is loaded on each machine, and the way in which an application is processed,
namely in parallel.
Parallel processing is the method of breaking down problems or work into smaller
components to be processed in parallel thus taking only a fraction of the time it would
take to run on a stand-alone PC.
The only drawback to this cost-effective way of computing is how can we effectively
design these systems to meet our performance needs? Can widely used operating
systems be used such as Windows? What software is available for users on this type
of machine and how do we run this software on other machines built using the same
technology? Can we use existing applications or do we need to develop new ones, if
so how? How can we ensure that each PC is doing its fair share of work, or is not
overloaded?
This document explores these issues from theory to practice, details a design
methodology and shows by experimental investigation that from a structured design
the speedup obtained with many PC’s can be known within bounds prior to
implementation.
To achieve our initial cost-effectiveness the Linux Operating system is used, however
Windows NT can be used if desired while still maintaining a competitive edge over
traditional super-computers. Additionally programming languages are available that
abstract from the system and free the programmer up from worrying about system
details.
Richard S. Morrison
B.E. (Computer Systems) Hons
MIEEE, MIEAust
February 2003
Cluster Computing
Page 5
Revision Version 2.3
Copyright © 2003
Contents
1.
INTRODUCTION....................................................................................................................................... 12
1.1. BRIEF HISTORY OF COMPUTING AND NETWORKING .............................................................................. 12
1.2. PARALLEL PROCESSING ......................................................................................................................... 12
1.3. MOTIVATION.......................................................................................................................................... 13
1.3.1.
Applications of Parallel Processing.............................................................................................. 14
2.
ARCHITECTURES.................................................................................................................................... 17
2.1. COMPUTER CLASSIFICATION SCHEMES .................................................................................................. 17
2.2. CLUSTER COMPUTING CLASSIFICATION SCHEMES ................................................................................. 21
2.3. BEOWULF............................................................................................................................................... 22
2.3.1.
History........................................................................................................................................... 22
2.3.2.
Overview ....................................................................................................................................... 23
2.3.3.
Classification................................................................................................................................. 24
2.4. NOW/COW........................................................................................................................................... 25
2.5. DISTRIBUTED VS. CENTRALIZED SYSTEMS ............................................................................................ 26
3.
SYSTEM DESIGN...................................................................................................................................... 28
3.1. PERFORMANCE REQUIREMENTS ............................................................................................................. 29
3.1.1.
The Need for Performance Evaluation.......................................................................................... 29
3.1.2.
Performance Indices of Parallel Computation.............................................................................. 30
3.1.3.
Theoretical Performance of Parallel Computers .......................................................................... 31
3.1.4.
Performance Analysis and Measurement...................................................................................... 36
3.1.5.
Practical Performance of Parallel Computers.............................................................................. 36
3.2. HARDWARE PLATFORMS ........................................................................................................................ 38
3.2.1.
CPU............................................................................................................................................... 38
3.2.2.
Symmetric Multiprocessing ........................................................................................................... 38
3.2.3.
Basic Network Architectures ......................................................................................................... 40
3.2.3.1.
Network Channel Bonding ........................................................................................................................42
3.2.4.
Node Interconnection Technologies.............................................................................................. 43
3.3. OPERATING SYSTEMS ............................................................................................................................ 44
3.3.1.
General.......................................................................................................................................... 44
3.3.2.
Towards Parallel Systems ............................................................................................................. 44
3.3.3.
Implementations ............................................................................................................................ 46
3.3.4.
Redhat Linux 7.2 ........................................................................................................................... 46
3.3.5.
Microsoft Windows 2000............................................................................................................... 50
3.3.6.
Sun Solaris .................................................................................................................................... 54
3.3.7.
Other ............................................................................................................................................. 54
3.4. MIDDLEWARE ........................................................................................................................................ 55
3.4.1.
Parallel Communications Libraries.............................................................................................. 56
3.4.1.1.
3.4.1.2.
3.4.2.
3.4.2.1.
3.4.2.2.
Page 6
PVM Overview..........................................................................................................................................56
MPI Overview ...........................................................................................................................................57
Application Development Packages.............................................................................................. 59
BSP............................................................................................................................................................61
ARCH........................................................................................................................................................61
Cluster Computing
Revision Version 2.3
4.
Copyright © 2003
SYSTEM INSTALLATION & TESTING................................................................................................ 63
4.1. BUILDING A BEOWULF ........................................................................................................................... 63
4.2. PERFORMANCE TESTING ........................................................................................................................ 66
4.2.1.
Beowulf Performance Suite ........................................................................................................... 66
4.2.2.
The Linpack Benchmark................................................................................................................ 67
4.3. SYSTEM ADMINISTRATION ..................................................................................................................... 68
4.3.1.
General.......................................................................................................................................... 68
4.3.2.
Mosixview...................................................................................................................................... 68
4.4. APPLICATIONS TESTING ......................................................................................................................... 69
4.4.1.
Persistence of Vision ..................................................................................................................... 69
5.
RESULTS .................................................................................................................................................... 73
5.1.
5.2.
SUMMARY OF NUMERICAL DATA .......................................................................................................... 73
RESULTS ANALYSIS ............................................................................................................................... 73
6.
CONCLUSION............................................................................................................................................ 76
7.
REFERENCES............................................................................................................................................ 78
8.
APPENDIX .................................................................................................................................................. 85
8.1. APPENDIX A – NODE INTERCONNECTION TECHNOLOGIES ..................................................................... 86
8.1.1.
Class 1 Network Hardware ........................................................................................................... 86
8.1.2.
Class 2 Network Hardware ........................................................................................................... 91
8.2. APPENDIX B – CHANNEL BONDING ....................................................................................................... 99
8.3. APPENDIX C – MPI IMPLEMENTATIONS............................................................................................... 105
8.3.1.
LAM............................................................................................................................................. 106
8.3.2.
MPICH ........................................................................................................................................ 107
8.4. APPENDIX D – POV-RAY .................................................................................................................... 108
8.4.1.
POV-Ray Benchmark .................................................................................................................. 108
8.4.2.
Alternate Bench Mark POV Files................................................................................................ 110
8.4.3.
POV-Ray Benchmark Performance data..................................................................................... 119
8.5. APPENDIX E – RAW RESULTS .............................................................................................................. 120
8.5.1.
Lam MPI Cluster boot-up and tear-down ................................................................................... 120
8.5.2.
POV-Ray ..................................................................................................................................... 121
Cluster Computing
Page 7
Revision Version 2.3
Copyright © 2003
Table of Figures
FIGURES
FIGURE 1-1 – AUTHOR COMPILED BEOWULF CLASS 1 HANDBOOKS ....................................................................... 3
FIGURE 1-1 – INTERACTION AMONG EXPERIMENT, THEORY AND COMPUTATION ................................................. 14
FIGURE 1-2 – ACSI WHITE AT THE LAWRENCE LIVERMORE NATIONAL LABORATORY ........................................ 16
FIGURE 2-1 – THE FLYNN-JOHNSON CLASSIFICATION OF COMPUTER SYSTEMS .................................................... 19
FIGURE 2-2 – ALTERNATE TAXONOMY OF PARALLEL & DISTRIBUTED COMPUTER SYSTEMS ............................... 19
FIGURE 2-3 – A PIPELINED PROCESSOR ................................................................................................................. 20
FIGURE 2-4 – SPACE-TIME DIAGRAM FOR A PIPELINE PROCESSOR........................................................................ 20
FIGURE 2-5 – SPACE-TIME DIAGRAM FOR A NON-PIPELINED PROCESSOR ............................................................. 21
FIGURE 2-6 – CLUSTER COMPUTER CLASSIFICATION SCHEME .............................................................................. 21
FIGURE 2-7 – CLUSTER COMPUTER ARCHITECTURE .............................................................................................. 22
FIGURE 2-8 – AVALON BEOWULF AT LANL ......................................................................................................... 23
FIGURE 2-9 – LOGO FROM THE BERKLEY NOW PROJECT ..................................................................................... 25
FIGURE 3-1 – LAYERED MODEL OF A CLUSTER COMPUTER .................................................................................. 28
FIGURE 3-2 – MAIN STEPS LEADING TO LOSS OF PARALLELISM ........................................................................... 29
FIGURE 3-3 – VARIOUS ESTIMATES OF AN N-PROCESSOR SPEEDUP ....................................................................... 35
FIGURE 3-4 – INTEL SMP SERVER BLOCK DIAGRAM ............................................................................................ 39
FIGURE 3-5 – SYSTEM COMPONENT RELATIONSHIPS IN THE LINUX OPERATING SYSTEM ..................................... 47
FIGURE 3-6 – LINUX HIERARCHICAL FILE SYSTEM STRUCTURE............................................................................ 48
FIGURE 3-7 – WINDOWS 2000 ARCHITECTURE...................................................................................................... 50
FIGURE 3-8 – WINDOWS 2000 CLUSTER SERVER TOPOLOGICAL DIAGRAM .......................................................... 51
FIGURE 3-9 – WINDOWS 2000 CLUSTER SERVER BLOCK DIAGRAM...................................................................... 52
FIGURE 3-10 – SUN CLUSTER STRUCTURE ............................................................................................................ 54
FIGURE 3-11 – THE PARALLEL PROGRAMMING MODEL EFFICIENCY ABSTRACTION TRADE-OFF .......................... 60
FIGURE 4-1 – LAYERED MODEL OF BEOWULF CLUSTER COMPUTER USED IN TESTING ........................................ 63
FIGURE 4-2 – TEST CLUSTER COMPUTER IN LAB B1/2122E.................................................................................. 66
FIGURE 4-3 – BPS RUNNING ON NODE1 OF THE TEST CLUSTER ............................................................................ 66
FIGURE 4-4 – MAIN WINDOW OF MOSIXVIEW CLUSTER MANAGEMENT SOFTWARE ............................................. 68
FIGURE 4-5 – RAY TRACING BENCH MARK TEST OUTPUT .................................................................................... 69
FIGURE 4-6 – SAMPLE OUTPUT OF POVRAY’S ABILITIES ....................................................................................... 70
FIGURE 4-7 – ALTERNATE POV-RAY TEST CASE 1............................................................................................... 70
FIGURE 4-8 – ALTERNATE POV-RAY TEST CASE II............................................................................................... 71
FIGURE 5-1 – TEST RESULTS & THEORETICAL PERFORMANCE COMPARISON ....................................................... 75
TABLES
TABLE 3-1 – TOPOLOGICAL PARAMETERS OF SELECTED INTERCONNECTION NETWORKS ..................................... 41
TABLE 5-1 – RENDERING TEST RESULTS ............................................................................................................... 73
TABLE 8-1 – LAM MPI 2 IMPLEMENTATION DETAILS ........................................................................................ 106
TABLE 8-2 – MPICH MPI 2 IMPLEMENTATION DETAILS .................................................................................... 107
TABLE 8-3 – POVBENCH POV-RAY PERFORMANCE BENCHMARK DATA EXTRACT ........................................ 119
Page 8
Cluster Computing
Revision Version 2.3
Copyright © 2003
Definitions and Acronyms
The following words and acronyms have been used throughout this document:
Word List
Definition
AI
Artificial Intelligence
API
Application Programming Interface
ATM
Asynchronous Transfer Mode
BSP
Bulk Synchronous Parallel
CAD
Computer Aided Design
CAI
Computer Assisted Instruction
CAM
Computer Aided Manufacturing
CAPERS
Cable Adapter for Parallel Execution and Rapid
Synchronization
CESDIS
Centre of Excellence in Space Data and Information
Sciences
CORBA
Common Object Request Broker Architecture
COTS
Commodity off the shelf
COW
Cluster of Workstations
CPU
Central Processing Unit
DEC
Digital Equipment Corporation (now Compaq)
DLM
Distributed Lock Manager
DMMP
Distributed Memory / Message Passing
DMSV
Distributed Memory / Shared Variables
DOE
US Department of Energy
ESS
Earth and Space Sciences Project – NASA
Full-Duplex
Simultaneous Bi-directional (transmitting and
receiving) communications over the same physical
communication link.
GFLOPS
1x109 FLoating Point OPerations per Second
GMMP
Global Memory / Message Passing
GMSV
Global Memory / Shared Variables
GSFC
Goddard Space Flight Centre
GUI
Graphical User Interface
HAL
Hardware Abstraction Layer
Half-Duplex
Bi-directional communications however transmitting
or receiving is mutually exclusive.
HPCC
High Performance Computer Community
Cluster Computing
Page 9
Revision Version 2.3
Page 10
Copyright © 2003
Word List
Definition
IEEE
Institution of Electrical and Electronics Engineers
IA32
Intel Architecture, 32 bit
IMPI
Interoperable MPI
LAM
Local Area Multicomputer
LAN
Local Area Network
LANL
Los Alamos National Laboratory
Linux
Unix style Operating system with open source code
arrangement. Linux is a trademark of Linus Torvalds.
LIPS
Logic Inferences Per Second
LWP
Light Weight Process
MFLOPS
1x106 FLoating Point OPerations per Second
MIPS
Million Instructions Per Second
MPI
Message Passing Interface Standard
MPP
Massively Parallel Processors
NASA
National Aeronautical & Space Administration
NFS
Networked File System
NIC
Network Interface Card
NOW
Network of Workstations
NPB
NASA Parallel Benchmark
OA
Office Automation
OS
Operating System
POSIX
Portable Operating System Interface Standard
uniX like
PVM
Parallel Virtual Machine
RED HAT
Is a registered trademark of Red Hat, Inc.
RFC
Request For Comments - Internet standard
RPC
Remote Procedure Call
SAN
System Area Network
SMP
Symmetric Multi Processing
SPP
Standard Parallel Port
SSI
Single System Image
Stream
Serial sequence of instructions or data
Telco
Telecommunications Company
TFLOPS
1x1012 FLoating Point OPerations per Second
ULT
User Level Thread
W2K
Microsoft Windows 2000
Cluster Computing
Revision Version 2.3
Copyright © 2003
Conventions used within this document
The following formatting, structure and comments are given below to assist with the reading
and interpretation of this document:
Advantages – Highlight a particular word, statement, or section of a paragraph.
Distributed Systems – Keyword within the document.
(For reference purposes) – Information provided to assist explanation, but only if required by
the reader.
Focus Box
Focus Boxes have been added to provide additional information such as supporting
fundamental concepts and related material of interest.
Cluster Computing
Page 11
Revision Version 2.3
Copyright © 2003
1. Introduction
1.1. Brief History of Computing and Networking
The use of computers within our society has developed from the very first usage in 1945 when
the modern computer era began, until about 1985 when computers were large and expensive.
At that time, even mini-computers normally cost tens of thousands of dollars. So as a result,
many organizations were limited to a handful of computers. These mini-computers were also
limited in the way Organizations could utilize them as they operated independently from each
other, as there was limited ability to interconnect them.
性能强的微处理器
高速的网络
During the mid 1980’s, however, two advances in technology began to change that situation.
The first was the development of powerful microprocessors. Initially these were only 8-bit, but
16, 32 and eventual 64 bit have followed. Many of these had the computing power of
mainframe systems (i.e. the larger systems) but at a fraction of the price.
The second development was the invention of the high speed LAN. Networking systems
allowed tens to hundreds of machines to be interconnected in such a way that small amounts of
information could be transferred between machines in only a few milli-seconds. As the speed
of networks increased larger amounts of data could be transferred.
The net result of these two technologies is the ability to construct ‘networks of machines’
composed of a large number of CPU’s interconnected via high-speed networks. These systems
are commonly known or referred to as Distributed Systems, in contrast to the centralized
systems consisting of a single CPU, local memory and I/O devices (such as mainframes that
have dominated prior to the development of low cost CPU’s and high speed networks). [5]
1.2. Parallel Processing
Parallel processing is the method of breaking large problems down into smaller constituent
components, tasks or calculations that are solvable in parallel. Parallel processing has emerged
as a key enabling technology in modern computing. The past several years have witnessed an
ever-increasing acceptance and adoption of parallel processing, both for high performance
scientific computing and for more “general purpose” applications. This has been a result of a
demand for higher performance, lower cost, and sustained productivity. The acceptance of
parallel processing has been facilitated by two major developments: massively parallel
processors (MPPs) and the widespread use of distributed computing. [2]
MPPs are now the most powerful computers in the world. These machines combine a few
hundred to a few thousand CPUs in a single large cabinet with hundreds of Giga-bytes of
memory.
MPPs offer enormous computational power and are used to solve computational Grand
Challenge problems such as global climate modelling and medicine design. As simulations
become more realistic, the computational power required to produce them grows rapidly. Thus,
researchers on the cutting edge turn to MPPs and parallel processing in order to get the most
computational power possible.
The second major development affecting scientific problem solving is distributed computing.
Distributed computing is a process whereby a set of computers connected by a network is used
collectively to solve a single large problem. As more and more organizations have high-speed
local area networks interconnecting many general-purpose workstations, the combined
Page 12
Cluster Computing
Revision Version 2.3
无与伦比的
Copyright © 2003
computational resources may exceed the power of a single high-performance computer. In
some cases, several MPPs have been combined using distributed computing to produce
unequalled computational power.
The most important factor in distributed computing is cost. Large MPPs typically cost more
than $10 million. In contrast, users see very little cost in running their problems on a local set
of existing computers. It is uncommon for distributed-computing users to realize the raw
computational power of a large MPP, but they are able to solve problems several times larger
than they could using one of their local computers.
A common architectural element between distributed computing and MPP’s is the notion of
message passing. In all parallel processing, data must be exchanged between cooperating
tasks. In this area of research, several paradigms have been experimentally evaluated including
shared memory, parallelising compilers, and message passing. The message-passing model has
become the paradigm of choice, from the perspective of the number and variety of
multiprocessors that support it, as well as in terms of applications, languages, and software
systems that use it.
This document focuses on an area of Distributed systems known as clustering, where the
many interconnected machines work collectively to process instructions and data. This area
can also be described as parallel systems (or thought of as using distributed systems for
parallel processing – as it encompasses both hardware, operating systems, middleware and
software applications).
As centralised machines dominate the low end of computing, for reference purposes section
2.5 outlines the benefits/drawbacks of distributed systems are compared with centralised
systems.
1.3. Motivation
Where do we obtain the drive or necessity to build computers with massively parallel
processing subsystems or large processing capabilities?
Fast and efficient computers are in high demand in many scientific, engineering, energy
resource, medical, artificial intelligence, basic research areas and now the corporate
environment.
Large-scale computations are often performed in these application areas and as such Parallel
processing computers are needed to meet these demands.
My research in this area has shown that the requirements of a parallelised computer can be
drawn from one of two areas and is derived from two separate requirements. These two areas
are listed below:
1.
Production Oriented Computing – This is where data is processed in real time
applications where timeliness, availability, reliability, and fault tolerance is a
necessity. This type of computing can require both serial and inherently or
parallelisable applications.
2.
Non-Production research computing – This is where a large amount of computer
power is required to work on mainly inherently parallel problems, or parallelisable
applications.
面向产品的计算
非生产型研究计算
This document concentrates on the non-production oriented research side of computing,
however the requirements of production oriented computing are detailed. My personal
motivation in this area relates to Plasma research and hence will need to be able to develop
algorithms to work with Clusters and MPPs.
Cluster Computing
Page 13
Revision Version 2.3
Copyright © 2003
1.3.1.
Applications of Parallel Processing
Large-scale scientific problem solving involves three interactive disciplines as shown in Figure
1-1 below:
Suggest & Test
Theory
Experimental
Generate Data
(Physicists, Engineers, Chemists,
Biologists)
Suggest & interpret
experiments
Model real processes,
suggest experiments,
analyse data, control
apparatus
Computational
Theoretical
(Computer Scientists, Digital
Engineers, Computational
Physicists)
(Mathematicians, Physicists,
Chemists, Logicians)
Provide equations,
interpret results
Accurate calculation, large-scale
calculations, suggest theory
Figure 1-1 – Interaction among Experiment, Theory and Computation
As shown above theoretical scientists develop mathematical models that computer engineers
solve numerically; the numerical results may then suggest new theories. Experimental science
provides data for computational science, and the latter can model processes that are hard to
approach in the laboratory. Using computer simulations has several advantages:
1.
Computer simulations are far cheaper and faster than physical experiments.
2.
Computers can solve a much wider range of problems than specific laboratory
equipment can.
3.
Computational approaches are only limited by computer speed and memory capacity,
while physical experiments have many practical constraints.
Presented below are five distinct areas in which parallel processing can be grouped according
to the required processing objective. Within each category several representative application
areas have been identified. Further detail on these application areas can be found in Hwang [6].
It is interesting to note that the majority of the top 500 supercomputers are running one of the
applications listed below. [29]
预测模型和模拟
Predictive Modelling and Simulations
Multidimensional modelling of the atmosphere, the earth environment, outer space, and the
world economy has become a major concern of world scientists. Predictive Modelling is done
through extensive computer simulation experiments, which often involve large-scale
computations to achieve the desired accuracy and turnaround time. Such numerical modelling
requires state-of-the-art computing at speeds approaching 1 GFLOPS and beyond.
Applications include:
Page 14
•
Numerical Weather Forecasting.
•
Semiconductor Simulation.
•
Oceanography.
•
Astrophysics (Modelling of Black holes and
Astronomical formations).
•
Sequencing of the human genome.
•
Socio-economic and Government use.
Cluster Computing
Revision Version 2.3
Copyright © 2003
工程设计与自动控制
Engineering Design and Automation
Fast supercomputers have been in high demand for solving many engineering design problems,
such as the finite-element analysis needed for structural designs and wind tunnel experiments
for aerodynamic studies. Industrial development also demands the use of computers to advance
automation, artificial intelligence, and remote sensing of earth resources.
Applications include:
•
Finite-element analysis.
•
Computational aerodynamics.
•
Remote Sensing Applications.
•
Artificial Intelligence and Automation – This areas requires parallel processing for
the following intelligence functions:
- CAD/CAM/CAI/OA
- Image Processing
- Intelligent Robotics
- Pattern Recognition
- Expert Computer Systems
- Computer Vision
- Knowledge Engineering
- Speech Understanding
- Machine inference
Energy Resources Exploration
能源探测
Energy affects the progress of the entire economy on a global basis. Computers can play an
important role in: the discovery of oil and gas, the management of their recovery, development
of workable plasma fusion energy and in ensuring nuclear reactor safety. Using computers in
the high-energy area results in less production costs and higher safety measures.
Applications include:
•
Seismic Exploration.
•
Plasma Fusion Power.
•
Reservoir Modelling.
•
Nuclear Reactor Safety.
Medical, Military and Basic Research
医学、军事与基础研究
In the medical area, fast computers are needed in computer-assisted tomography (CAT scan),
imaging, artificial heart design, liver diagnosis, brain damage estimation and genetic
engineering studies. Military defence need to use supercomputers for weapon design, effects
simulation, and other electronic warfare. Almost all basic research areas demand fast
computers to advance their studies.
Applications include:
• Medical Imaging
•
Polymer Chemistry.
• Quantum Mechanics problems.
•
Nuclear Weapon Design.
Visualization
可视化
Films such as The Matrix, Titanic, and Toy Story all made extensive use of cluster computers
to generate the huge amount of imagery that was required to make these films. The cost of
generating physical scenery for fantasy realms and historic recreations can be replaced by
rendering on computers, not to mention the ability to violate the laws of physics.
Applications include:
Cluster Computing
•
Computer-generated graphics, films and animations.
•
Data Visualization.
Page 15
Revision Version 2.3
Copyright © 2003
World’s Fastest Computers
The power of the last decade's supercomputers are now available in affordable desktop
systems. However, the demand for ever-increasing computational power has not abated. For
example, the US Department of Energy (DOE) established a program in 1996 called the
Accelerated Strategic Computing Initiative (ASCI). The goal of ASCI is to enable 100 trillion
floating point operations per second (teraflop/s or TFLOPS) sustained performance on key
simulation codes by the year 2005. The DOE's aggressive schedule is due to the need to
replace a test-based process for maintaining the nuclear stockpile with a simulation-based
process. The change from test-based to simulation-based processes must be completed before
the retirement of the aging population of nuclear scientists who can verify the conversion. At
the same time industries from oil-and-gas exploration to aeronautics to pharmaceuticals are
beginning to discover the enormous power inherent in high fidelity, three-dimensional
simulations. [28]
In previous years, the world’s most powerful computational machine was the ASCI White
[29], an IBM-designed RS/6000 SP system that marked a breakthrough in computing at 12.3
teraflops. ASCI White uses MPP technology and is powered by 8,192 copper microprocessors,
and contains 6 Tera bytes of memory with more than 160 TB of disk storage capacity.
Figure 1-2 – ACSI White at the Lawrence Livermore National Laboratory
IBM is currently in the process of building a new supercomputer for the DOE called Blue
Gene/L. The computer is due for completion by 2005 and will operate at about 200 teraflops
which has in previous years been larger than the total computing power of the top 500
supercomputers.
While the upper echelon of supercomputing relies on MPP technology, which will remain
dominant for a long time, it is interesting to note that MPP and clusters are based around the
same information theory and use familiar programming paradigms as discussed in this
document, namely the message passing standard MPI.
As a point of comparison, at the time of writing, 63% of the world’s top 500 supercomputers
use MPP, 7% use SMP and 6% are cluster based. [29]
Page 16
Cluster Computing
Revision Version 2.3
Copyright © 2003
2. Architectures
2.1. Computer Classification Schemes
In general, digital computers may be classified into four categories, according to the
multiplicity of instruction and data streams. Michael J. Flynn introduced this scheme or
taxonomy for classifying computer organizations in 1966.
The essential element of the computing process is the execution of a sequence of instructions
on a set of data.
Computer organizations are characterized by the multiplicity of the hardware provided to
service the instruction and data streams. Listed below are Flynn’s four machine organizations:
•
SISD – Single Instruction stream / Single Data stream
•
SIMD – Single Instruction stream / Multiple Data stream
•
MISD – Multiple Instruction stream / Single Data stream (No real application)
•
MIMD – Multiple Instruction stream / Multiple Data stream
Distributed systems can be classified by the MIMD organization but as each node can be
classified as SISD both these schemes are presented here:
SISD – This organization represents most serial computers available today. Instructions are
executed sequentially but may be overlapped in their execution stages (using pipelining – refer
to focus box below). Most uniprocessor systems are pipelined. A SISD computer may have
more than one functional unit in it. All the functional units are under the supervision of one
control unit.
MIMD – Most multiprocessor systems and multiple computer systems can be classified in this
category. An intrinsic MIMD computer implies interactions among the n processors because
all memory streams are derived from the same data space shared by all processors.
MIMD can be summarised as follows:
•
Each processor runs its own instruction sequence.
•
Each processor works on a different part of the problem.
•
Each processor communicates data to other parts.
•
Processors may have to wait for other processors or for access to data.
If the n data streams were derived from disjointed subspaces of the shared memories, then we
would have the so-called multiple SISD (MSISD) operation, which is nothing but a set of n
independent SISD uniprocessor systems.
An intrinsic MIMD computer is tightly coupled if the degree of interactions among the
processors is high. If not we consider them loosely coupled.
An example of the two:
Loosely Coupled multiprocessor systems do not generally encounter the degree of memory
conflicts experienced by tightly coupled systems. In such systems, each processor has a set of
input-output devices and a large local memory where it accesses most of the instructions and
data. We refer to the processor, its local memory and I/O interfaces as a computer module.
Cluster Computing
Page 17
Revision Version 2.3
Copyright © 2003
Processes that execute on different computer modules communicate by exchanging messages
through a message transfer system. (MTS). The degree of coupling in such a systems is very
loose.
The determinant factor of the degree of coupling is the communication topology of the
associated message transfer system. Loosely coupled systems are usually efficient when the
interactions between tasks are minimal.
However due to the large variability of inter-reference times, the throughput of the hierarchical
loosely coupled microprocessor may be too low for some applications that require fast
response times. If high-speed or real time processing is required tightly coupled systems may
be used.
An example of a loosely coupled multiprocessor machine is the Beowulf architecture, which is
the subject of this document. Each processor has its own machine (a separate PC) connected
through an Ethernet channel.
Tightly Coupled multiprocessor systems generally benefit over loosely coupled systems in
performance however cost significantly more to achieve this performance.
Tightly coupled multiprocessors communicate through a shared main memory. Hence the rate
at which processors can communicate from one processor to the other is on the order of the
bandwidth of the memory. Tightly coupled systems can tolerate a higher degree of interactions
between tasks without significant deterioration in performance. [6]
An example of a tightly coupled multiprocessor machine is a TransputerTM or a machine where
two or more processors share the same memory such as with SMP (Symmetric
Multiprocessing).
As the MIMD category includes a wide class of computers, in 1988 E. E. Johnson [9] proposed
a further classification of such machines based on their memory structure (global or distributed)
and the mechanism used for communications/synchronisation (shared variables or message
passing). These are presented below:
•
GMSV – Global Memory / Shared Variables.
•
GMMP – Global Memory / Message Passing (No real application)
•
DMSV – Distributed Memory / Shared Variables.
•
DMMP – Distributed Memory / Message Passing.
It can be seen from the above classification that two of these architectures; DMSV and DMMP
are loosely coupled and the remaining two; GMSV and GMMP are tightly coupled. As
message passing is a key component to the Beowulf cluster technology, its architecture is best
described or defined by the DMMP classification.
This leads us to the Flynn-Johnson classification of computer systems:
Page 18
Cluster Computing
Revision Version 2.3
Copyright © 2003
Data Stream(s)
Multiple
SIMD
SISD
GMSV
MISD
GMMP
MIMD
DMSV
DMMP
Distributed Global
Memory
Single
Multiple
Instruction Stream(s)
Single
Shared Variables
Message Passing
Communication / Syncronisation
Figure 2-1 – The Flynn-Johnson Classification of Computer Systems
细目分类
An alternate breakdown of the MIMD category proposed by Johnson is detailed by
Tanenbaum and is shown in Figure 2-2 below. [5]
MIMD
Parallel &
Distributed
Computers
Multiprocessors
(Shared Memory)
Mulitcomputers
(Private Memory)
BUS
SWITCHED
BUS
SWITCHED
Sequent,
Encore
Ultracomputer,
RP3
Workstations on
a LAN
Hypercube,
Transputer
Tightly Coupled
Loosely Coupled
Figure 2-2 – Alternate Taxonomy of Parallel & Distributed Computer Systems
Tannebaum’s taxonomy differs from the Flynn-Johnson model as it classifies a computer
system based on the architecture of the interconnection network where the Flynn-Johnson
model differentiates the programming model of the system. Both models are valid and present
the similar information, however the Flynn-Johnson model is more descriptive as it describes
the logical connections between processors rather than the physical connections and is thus
less hardware specific and therefore is a better classification model.
As such further discussion is based of the Flynn-Johnson Classification model.
Cluster Computing
Page 19
Revision Version 2.3
Copyright © 2003
Pipelining
As stated above all SISD computers generally use pipelining. This is the case with all INTEL
40x86 and later CPU generations. As these processors will make up the cluster computer that
will be tested the concept is explained below.
时间并行性
A pipelined computer performs overlapped computations to exploit temporal parallelism.
Normally the process of executing an instruction in a digital computer involves four major
steps:
1.
Instruction Fetch (IF) from main memory.
2.
Instruction Decoding (ID) identifying the operation to be performed.
3.
Operand Fetch (OF) if needed in the execution.
4.
Execution (EX) of the decoded arithmetic logic operation.
In a non-pipelined computer the four steps must be completed before the next instruction can
be issued.
S1
S2
S3
S4
IF
ID
OF
Stages
EX
Figure 2-3 – A Pipelined Processor
In a pipelined computer, successive instructions are executed in an overlapped fashion, as
illustrated in Figure 2-4. Four pipeline stages, IF, ID, OF and EX, are arranged into a linear
cascade. The two space-time diagrams as depicted in Figure 2-4 and Figure 2-5 show the
difference between overlapped instruction execution and sequentially non-overlapped
execution.
This explanation depicts an Instruction Pipeline only – integer pipelines are also common with
the INTEL Pentium containing two of them, potentially allowing two instructions to execute at
the same time.
Pipeline Stages
Complete Instruction Cycles
I1
I2
I3
I4
I1
I2
I3
I4
I5
I1
I2
I3
I4
I5
I2
I3
I4
I5
EX
OF
ID
IF
I1
1
2
3
4
5
6
I5
7
8
9
Time (Pipeline Stages )
Figure 2-4 – Space-Time Diagram for a Pipeline Processor
PTO
Page 20
Cluster Computing
Revision Version 2.3
Copyright © 2003
Pipelining (cont).
Stages
Complete Instruction Cycles
I1
EX
I1
OF
I3
I2
I1
ID
IF
I2
I3
I2
I1
2
3
4
6
7
8
I5
I4
I3
5
I5
I4
9
10
I5
I4
I3
I2
1
I4
11
12
13
I5
14
15
16
17
18
19
20
21
Time
Figure 2-5 – Space-Time Diagram for a Non-pipelined Processor
Theoretically a k-stage linear pipeline processor could be at most k times faster. However due to memory
conflicts, data dependency, branch and interrupts, this speedup may not be achieved for out-of-sequence
computations. [6] [7]
2.2. Cluster Computing Classification Schemes
Whilst the Flynn-Johnson DMMP model represents a number of different computer
implementations, DMMP is also is a high level description for the area of computing that this
document focuses on, which is the area of Cluster Computing.
Should we which to define DMMP machines further than the Flynn-Johnson model by way of
classification of architectures, we would be defining actual implementation specific machines.
One such classification is proposed by Thomas L. Sterling et al [12]:
DMMP
Cluster Computing
Pile of PCs
NOW/COW
WS Farms/
Cycle
Harvesting
Beowulf
NT-PC Clusters
DSHMEMNUMA
Figure 2-6 – Cluster Computer Classification Scheme
This document explicitly focuses on the Beowulf Cluster Computer, but does look at NOW
(Network of Workstations) and COW (Cluster of Workstations) as well as NT-PC Clusters all
of which are similar.
Cluster Computing
Page 21
Revision Version 2.3
Copyright © 2003
A cluster computer is a type of parallel or distributed processing system, which consists of a
collection of interconnected stand-alone computers working together as a single integrated
resource. The typical architecture of a cluster is shown in Figure 1-1 below. [31]
Figure 2-7 – Cluster Computer Architecture
The following are some prominent components of cluster computers:
•
Multiple high performance computers (PCs, Workstations, SMP servers).
•
Operating systems (Layered or Micro-Kernel based).
•
High Performance Networks (Switched network fabrics, Gigabit Ethernet).
•
Cluster Middleware (Single System Image, and System Availability Infrastructure).
•
Parallel Programming Environments.
•
Applications.
2.3. Beowulf
2.3.1.
类型、流派
History
In the summer of 1994 Thomas Sterling and Don Becker, working at CESDIS under the
sponsorship of the ESS project, built a cluster computer consisting of 16 Intel DX4 processors
connected by channel bonded Ethernet. They called their machine Beowulf. The machine was
an instant success and their idea of providing COTS (Commodity off the shelf) based systems
to satisfy specific computational requirements quickly spread through NASA and into the
academic and research communities. The development effort for this first machine quickly
grew into what we now call the Beowulf Project. A non-technical measure of success of this
initial project is the observation that researchers within the High Performance Computing
community now refer to such machines as "Beowulf Class Cluster Computers". That is,
Beowulf clusters are now recognized as genre within the HPC community. [4.2]
Thomas Sterling named the project ‘Beowulf’ after an old English tale of a legendary sixthcentury hero from a distant realm who freed the Danes of Heorot by destroying the oppressive
monster Grendel. As a metaphor, "Beowulf" was applied to this new strategy in high
performance computing that exploits mass-market technologies to overcome the oppressive 难以忍受的
costs in time and money of supercomputing, thus freeing scientists, engineers, and others to
devote themselves to their respective disciplines. Beowulf, both in myth and reality, challenges
and conquers a dominant obstacle, in their respective domains, thus opening the way to future
achievement. [12]
Page 22
Cluster Computing
Revision Version 2.3
2.3.2.
Copyright © 2003
Overview
A Beowulf is distinct from other cluster computers as each node (PC) does not have keyboards,
mice, video cards or monitors. For example, shown below is the LANL Beowulf machine
named Avalon. Avalon was built in 1998 using the DEC Alpha chip, resulting in very high
performance [24].
Figure 2-8 – Avalon Beowulf at LANL
Beowulf is a machine usually dedicated to parallel computing, and optimised for this purpose.
It gives a better price/performance ratio than other cluster computers built from Commodity
Components and runs mainly software that is available at no cost. Beowulf has also more
single system image features that assists users to utilize the Beowulf cluster as a single
computing workstation.
Beowulf utilizes the client/server model of computing in its architecture as does many
distributed systems (with the noted exception of peer-to-peer). All access to the client nodes is
done via remote connections from the server node, dedicated console node or a serial console.
As there is no need for client nodes to access machines outside the cluster, nor for machines
outside the cluster to access client nodes directly, it is common practice for the client nodes to
use private IP addresses such as the following IP address ranges:
10.0.0.0/8 or
192.168.0.0/16
The Internet Assigned Numbers Authority (IANA) has reserved these IP address ranges for
private Internets, as discussed in RFC 1918 [13].
Usually the only machine that is connected to an external network is the server node. This is
done using a second network card in the server node itself. The most common ways of using
the system is to access the server's console directly, or either telnet or remote login to the
server node from a personal workstation. Once on the server node, users are able to edit and
compile source-code, and also spawn jobs on all nodes in the cluster. [1]
Cluster Computing
Page 23
Revision Version 2.3
2.3.3.
Copyright © 2003
Classification
Beowulf systems can be constructed from a variety of distinct hardware and software
components. In order to increase the performance of Beowulf clusters some non-commodity
components can be employed. In order to account for the different types of hardware systems
and to simplify references about Beowulf systems architecture, Radjewski et al [1] proposed
the following simple classification scheme:
CLASS I BEOWULF:
Beowulf Class I machines are built entirely from commodity "off-the-shelf" parts. A Class I
Beowulf is a machine that can be assembled from parts found in at least 3 nationally/globally
circulated advertising catalogues.
The advantages of a Class I system are:
•
Hardware is available form multiple sources (low prices, easy maintenance)
•
No reliance on a single hardware vendor.
•
Driver support from Linux commodity.
•
Usually based on standards (SCSI, EIDE, PCI, Ethernet, etc).
The disadvantages of a Class 1 system are:
•
Best performance is attained using CLASS II hardware.
CLASS II BEOWULF
A Class II Beowulf is simply any machine that does not pass the ease of availability
certification test.
This class of Beowulf generally delivers higher performance at a cost penalty.
The advantages of a CLASS II system are:
•
Performance is generally better than class one depending on the design and
configuration (i.e. the use of a 1.2GBps LAN will only improve performance for
highly coupled applications).
The disadvantages of a CLASS II system are:
•
Driver support may vary.
•
Reliance on single hardware vendor for parts and pricing.
•
Generally more expensive than Class I systems.
Either class is not necessarily better than the other. It depends on the cost/performance
requirements of the system that is required.
The classification system is only intended to assist in the discussion and specification of
Beowulf systems.
This document focuses on Class I Beowulf systems as parts are highly available to construct
clusters.
Page 24
Cluster Computing
Revision Version 2.3
Copyright © 2003
2.4. NOW/COW
Networks of Workstations (NOW) and Clusters of Workstations (COW) differ physically from
Beowulf, as they are essentially complete PCs connected via a network. In most cases COWs
are used for parallel computations at night, and over weekends when people are not actually
using the workstations for every day work.
Another instance of this type of network is when people are using their PC but the central
server is using a portion of their overall processing power and aggregating this across the
network, thus utilising idle CPU cycles. In this instance programming a NOW is an attempt to
harvest unused cycles on each workstation. Programming in this environment requires
algorithms that are extremely tolerant of load balancing problems and large communication
latency. Any program that runs on a NOW will run at least as well on a cluster. [1] [4.2]
Figure 2-9 below is the logo from the Berkley NOW project, that demonstrates the principle of
building parallel processing computers from COTS components. [26]
Figure 2-9 – Logo from the Berkley NOW Project
Previous research on cluster computing indicated that there are strong limits on the
applications that will perform effectively on a COW or NOW. Beowulf improves on this
philosophy by the ability to make the following assumptions which improves performance
and/or the functionality:
Cluster Computing
•
Dedicated processors.
•
Private networks.
•
High performance networks.
•
Customisable system software.
•
Systems Software can be imaged, ‘cloned’ or bootstrapped over the network.
Page 25
Revision Version 2.3
Copyright © 2003
2.5. Distributed vs. Centralized Systems
Distributed Systems have many advantages and are discussed here to outline the motivations
of this area of work. Each advantage listed below is generally applicable to Beowulf clusters as
they are encompassed by distributed systems.
Although distributed systems have their strengths, they also have their weakness, the majority
of which are also outlined below. Despite these problems, it is considered that the advantages
outweigh the disadvantages. [5]
ADVANTAGES
高性能、高吞吐
高可用性
High Performance, High Throughput – With multiple interconnected systems, high
performance processing is attainable through the use of clustering where individual
machines work as one, or appear as one transparently to the user to do work.
High Availability – With multiple interconnected systems, the loss of any one system
should have only minimal impact. Key systems and components can be replicated so that
a backup system can take up the load after a failure.
高可靠性
High Reliability – Should one machine fail, the system as a whole will not fail.
经济成本
Economies of Scale – The real driving force behind the trend towards decentralization is
economic. Distributed systems have a better price performance ratio.
可扩展性、可扩充性
Expandability & Scalability – When additional computer power is required, more CPU’s
can be added onto the network and the software ‘cloned’ or copied from another machine
in a simple process. With centralised systems, upgrading a mainframe machine is cost
prohibitive as well as complex.
微微秒
技术优势
适合的应用
地理位置分布
并行
Technology – Today Clusters can be built out of a thousand CPUs yielding 20,000 MIPS,
a comparable centralised system cannot currently be found. Further, to build a system of
similar processing power would require a CPU that could execute an instruction in
0.005nsec (50 picosecond). From Einstein’s theory of relativity we know that the speed of
electrons have an upper limit of the speed of light (practically much lower than this due to
resistance and other factors). This places an implication on the system to be a 1.5cm cube
as we can only cover 1.5cm in 50picosec. Whilst current manufacturing processes will
advance and produce faster and faster systems, parallelisation is the more cost effective
alternative as we further get closer to the physical signal limit of the wavelength of light.
Considering this, the sooner we embrace SMP and Cluster technology the sooner we will
really see speed enhancements in our computer systems.
Suits Applications – Some applications are inherently distributed in nature, such a
geographically diverse operations or suit multiple processors working in parallel such as
inherently parallel problems. Examples of these:
Geographically distributed: such as a Factory where Robots and machines are
distributed along the assembly line. It is considered optimal to have individual local
computers to control and manage the immediate functions of each robot. Each
computer is interconnected via a network where the overall coordination, control and
statistics monitoring can be remotely administered.
Parallel: such as a multi-threaded server application, running on a multi-tasking
operating system, where each spawned thread can be allocated its own processor until
such time as multiple threads are running on each processor. In this scenario the
server will have a faster response time, as work can be done in parallel.
Page 26
Cluster Computing
Revision Version 2.3
Copyright © 2003
DISADVANTAGES
有用的软件有限
网络通信不可靠
安全
Cluster Computing
Available Software is Limited – Little software is commercially available that offers
functionality, reliability and can truly exploit a distributed environment. Additionally it
can be quite expensive to develop new software applications for this architecture.
Communications Network can Introduce Unreliability – It can potentially lose
messages, become saturated, or too bandwidth limited to effectively serve the machines
connected to it. To rectify these problems requires additional software and
communications protocols and can require an expensive investment in new hardware.
Introduces Security Problems – The ease of sharing data can be noted as an advantage
of a distributed system. However as people and other systems can conveniently access
data from any node in the system, it is equally possible to be able to access data that they
have no business accessing. Security concerns in a distributed system are more far ranging
that a centralised system as there is multiple points to, and of, attack and with distributed
data systems, a greater risk of internal accidental data retrieval/deletion. This situation is
possible to counter with rigorous lock-down procedures if followed by the
network/systems integrators and administrators.
Page 27
Revision Version 2.3
Copyright © 2003
3. System Design
As previously stated; Fast, efficient, highly available and reliable computers are in high
demand in many scientific, engineering, energy resource, medical military, artificial
intelligence, research and now this requirement is seen in the corporate environment.
Whilst the requirements of each specific industry is different the key requirement is the same:
Design a cost effective super-computer.
With this chief requirement in mind, it is quite impossible to look past clustering as the
solution.
This section approaches the system design requirements of a Beowulf cluster computer using a
bottom-up design methodology as follows:
1.
Performance Requirements – What performance is desired? What Performance is
required? To effectively design a cluster computer system we need to know what
application(s) will be running on the cluster and what is the intended purpose. This
section reviews the theory behind parallel processing performance issues.
2.
Hardware – What Hardware will be required or is available to build the Cluster
Computer.
3.
Operating Systems – What operating systems are available for Cluster Computing
and what are their benefits.
4.
Middleware – Between the Parallel application and the operating system we need a
way of load balancing, and managing the system, users and processes, unless this is
incorporated in to the Operating System.
5.
Applications – What Software is available for Cluster Computing and how can
applications be developed?
性能需求
硬件
操作系统
中间件
应用
To assist in presentation of the material, this section has been structured following the
bottom-up approach using the basic layered model of a cluster computer as shown in
Figure 3-1 below. The hardware is shown at the bottom (i.e. discussed first) and the
successive building blocks (dependencies) are discussed in later headings. This structure
has been paid particular attention in the section on middleware as many packages can
make up this layer.
Figure 3-1 – Layered Model of a Cluster Computer
Page 28
Cluster Computing
Revision Version 2.3
Copyright © 2003
3.1. Performance Requirements
3.1.1.
The Need for Performance Evaluation
One of the most important issues in parallel processing is how to effectively utilize parallel
computers due to their inherent complexity. It is estimated that many modern supercomputers
and parallel processors deliver only 10% or less of their peak performance potential in a
variety of applications. Yet high performance is the very reason why super-computers are
built.
The causes of performance degradation are many. Performance losses occur because of
mismatches among:
1.
Hardware – Idle processors due to conflicts over memory access & communications
paths.
2.
Operating System – Inefficient internal scheduler, file systems and memory
allocation/de-allocation.
3.
Middleware – Inefficient distribution and coordination of tasks, high inter-processor
communications latency due to inefficient middleware.
4.
Applications – Inefficient algorithms that do not exploit the natural concurrency of a
problem.
The main steps leading to loss of parallelism, ranging from problem domain to hardware, are
shown in Figure 3-2 below.
m
er
em
a re
p il
rith
yst
m
o
dw
S
r
o
g
l
a
g
C
A
H
&
tin
ge
era
a
p
u
O
ng
La
m
b le
P ro
Figure 3-2 – Main Steps Leading to Loss of Parallelism
The degree of parallelism is the number of independent operations that may be performed in
parallel. In complex systems, mismatches occur even among software and hardware modules.
For example, the communications network bandwidth may not correspond to the processor
speed or that of the memory.
Mapping applications to parallel computers and balancing processor loads is indeed a very
difficult task.
Cluster Computing
Page 29
Revision Version 2.3
Copyright © 2003
Acquisition of the following knowledge motivates performance evaluation:
1.
Evaluation and comparison of new and existing parallel computers.
2.
Increasing efficiency in processor utilization by matching algorithms and
architectures.
3.
Acquisition of information for the design of new computers systems.
In general, computer performance can be studied by analytical means, by measurements and
analysis. However with respect to parallel computers due to the complexity, analytical models
are extremely complex and too many simplifying assumptions need to be made such that the
end result does not accurately represent the operation of a parallel computer.
At present empirical results are the only methods that can be relied upon to assess the
performance of parallel computers.
Performance benchmarking is another way to assess the performance of a particular parallel
system. Benchmarking programs for Beowulf systems such as the LINPACK or the NASA
developed NPB package are methods to compare Beowulf systems. Whilst benchmarks are
useful, such work does not address the problem of how to tune the architecture, hardware,
system software, and application algorithms to improve the performance. We need to gain an
understanding of why a machine is performing the way it is through analysing how a machine
performs across a wide range of benchmarks. [14]
3.1.2.
Performance Indices of Parallel Computation
Within the research and development work that has been conducted in the field of parallel
systems, some performance indices have been defined to measure the performance of parallel
computation. Due to the level of complexity involved, no single measure of performance can
give a truly accurate measure of a computer systems performance. Different indices are needed
to measure different aspect. Some indices for global measurements follow [9] [14]:
1.
Execution rate – The execution rate measures the machine output per unit of time.
Depending on how machine output is defined, several execution rate measurements
can be used. The concept of instructions per second, measured in MIPS (million
instructions per second), is often used. While this measurement is meaningful for
uniprocessors and multiprocessors it is inappropriate for SIMD machines, in which
one instruction operates on a large number of operands. Additionally it does not
differentiate between true results and overhead instructions. A good measure for
arithmetic operations is Megaflops (Mflops). While this measure is appropriate for
many numeric applications, it is not very useful for AI programs. A measure that can
be used for logic programs and which can also be used for AI is LIPS (logic
inferences per second).
2.
Speedup (Sp) – The speedup factor of a parallel computation using p processors is
defined as the ratio:
执行速度
加速度
Sp =
T1
Tp
Where T1 is the time taken to perform the computation on one processor and
Tp is the time taken to perform the same computation on p processors.
Hence Sp is the ratio of the sequential processing time to the parallel processing time.
Normally the speedup factor is normally less than the number of processors because
of the time lost to synchronisation, communication time, and other overheads required
by the parallel computation:
1≤ Sp ≤ p
However there are cases where this does not apply. Refer to Super-Linear Speedups
focus box below for details.
Page 30
Cluster Computing
Revision Version 2.3
Copyright © 2003
3.
效率
Efficiency (Ep) – The efficiency of a parallel computation is defined as the ratio
between the speedup factor and the number of processors:
Ep =
Sp
P
=
T1
pT p
Efficiency is a measure of the cost-effectiveness of computations.
4.
Redundancy (Rp) – The redundancy of a parallel computation is the ratio between
the total number of operations Op executed in performing some computation with p
processors and the number of operations O1 required to execute the same computation
with a uniprocessor:
冗余率
Rp =
Op
O1
Rp is related to the time lost because of overhead, and is always larger than 1.
5.
Utilization (Up) – The utilization factor is the ratio between the actual number of
operations Op and the number of operations that could be performed with p processors
in Tp time units:
利用率
Up =
6.
Op
pT p
Quality (Qp) – The Quality factor of a parallel computation using p processors is
defined as the equation:
3
质量
T1
Qp =
2
pTP OP
Super-Linear Speedups
Superlinear speedups have been reported by Molavan [14] for non-deterministic AI
computations, especially search operations. With respect to parallel processing with search
operations, some paths leading to wrong results may be eliminated earlier than with sequential
processing. Thus by avoiding some unnecessary computations, the speedup may increase by
more than the number of processors.
One example of an algorithm that exhibits a super-linear speedup is A* - a widely known
branch-and-bound algorithm used to solve combinatorial optimisation problems. [3.4.2.2][34]
3.1.3.
Theoretical Performance of Parallel Computers
In this section, some basic mathematical models of parallel computation are presented. They
are useful for understanding the limits of parallel computation.
The ideal speedup that can be achieved by a parallel computer with n identical processors
working concurrently on a single problem is at most n times faster than a single processor. In
practice, the speedup is much less, due to the many factors as outlined above.
The lower bound log2n is known as Minsky’s conjecture, that the speedup is proportional to 猜想
the logarithm of the number n of processors. This conjecture has its roots in analysis of data
access conflicts assuming random distribution of addresses. These conflicts will slow all
processes down to the point that quadrupling the number of processors only doubles the
performance. However, data access patterns in real applications are far from random. Most
applications have a good percentage of data access regularity and locality (refer to the locality
of reference focus box below) that help improve the performance.
Cluster Computing
Page 31
Revision Version 2.3
Copyright © 2003
悲观
This law is considered to be pessimistic or the lower bound of the speedup. Depending on the
application, real speed-up can range from log2n to n. However, with respect to the upper bound
n, in practice the formula (and accompanying derivation) below tends to model the
performance of the upper bound more accurately.
Consider a computing problem, which can be executed by a uniprocessor in:
Unit time T1 =1
Let fi be the probability of assigning the same problem to i processors
Each processor is working equally with average load di =1/i per processor
Assume equal probability of each operating mode using i processors i.e. fi = 1/n, for
n operating modes: I = 1, 2,…..n
The average time required to solve the problem on an n-processor system is given
below, where the summation represents n operating modes:
n
1
n
∑i
i =1
n
Tn = ∑ f i .d i =
i =1
Equation 3-1
The average speedup S is obtained as the ratio of T1 =1 to Tn; that is,
S=
T1
n
n
= n ≤
1 ln n
Tn
∑
i =1 i
Equation 3-2
The Weighted Harmonic Mean Principle relates the execution rate to a weighted harmonic
mean. Let Tn be the time required to compute n tasks of k different types. Each type consists of
ni tasks requiring ti seconds each, such that:
k
Tn = ∑ ni .ti
1
And
k
n = ∑ ni
1
By definition, the execution rate R is the number of events or operations in unit time, so:
Rn =
n
=
Tn
n
n
∑ n .t
i
i
1
Let us define fi as the fraction of results generated at rate Ri, and ti = 1/Ri as the time required
to generate a result. Then:
Rn =
1
=
k
1
k
∑ f .t ∑ f
i
1
i
i
Equation 3-3
/ Ri
1
Where
fi =ni/n
Page 32
Cluster Computing
Revision Version 2.3
Copyright © 2003
∑f
i
=1
i
Ri = 1/ti
Equation 3-3 represents the basic principle in computer architecture some times loosely
referred to as a bottleneck, or the weighted harmonic mean principle. The importance of this
principle and Equation 3-3 is the following:
If a single rate is out of balance (i.e. lower than the others), then this will dominate the
overall performance of the machine.
This is the case for designing general-purpose Beowulf cluster client-nodes with
homogenous hardware. This is as opposed to the use of heterogeneous collections of nodes
communicating over various different networks. A heterogeneous network and node base are
systems that are out of the range of mathematical modelling or determining the performance
prior to implementation.
Amdahl introduced Amdahl’s law in 1967. This law is a particular case of the weighted
harmonic mean principle. Two rates are considered:
1.
The high, or parallel execution rate RH
2.
The low, or scalar execution rate RL
If f denotes the fraction of results generated at the high rate RH and 1- f is the fraction
generated at the low rate, then the Equation 3-3 becomes:
Rf =
1
f / RH + (1 − f ) / RL )
Equation 3-4
This formula is known as Amdahl’s law. It is useful for analysing system performance that
results from two individual rates of execution, such as vector or scalar operations or parallel or
serial operations. It is also useful for analysing a complete parallel system in which one rate is
out of balance with the others. For example, the low rate may be caused by I/O or
communication operations, and the high rate may be caused by vector, memory, or other
operations.
Amdahl’s law can also be applied to the execution of a particular program. From the law we
see that a small fraction f of inherently sequential or unparalisable computation severely limits
the speedup that can be achieved with p processors. Consider a unit time task, for which the
fraction f is unparalisable (so it takes the same time f on both parallel and sequential
machines) and the remaining 1-f is fully paralisable [so it runs in time (1-f)/p on a p-processor
machine].
Of note is the fact that with Amdahl’s formula when f=0 this represents the ideal case of and n
times speedup.
Hence to plot these results will give an indicative idea of what the performance a Beowulf
cluster will produce depending on the number of processors that are used. In practice this will
vary, however, if a good communications network is chosen the speedup as shown will be
achieved. Shown below are three views from the same data derived from the principles above.
The three views show; the over all trend and, a low number of processors.
Cluster Computing
Page 33
Revision Version 2.3
Copyright © 2003
512
448
Speedup
384
320
256
192
128
64
0
0
64
128
192
256
320
384
448
512
28
32
Number of Processors
32
28
Speedup
24
20
16
12
8
4
0
0
4
8
12
16
20
24
Number of Processors
Page 34
Cluster Computing
Revision Version 2.3
Copyright © 2003
8
Speedup
6
4
Log2n
n/lnn
Amdahls f = 0.05
2
Amdahls f = 0.1
Ideal
0
0
2
4
6
8
Number of Processors
Figure 3-3 – Various Estimates of an n-processor Speedup
From the above analysis and plot it can be seen why typical commercial multiprocessor
systems consist of only two or four processors. Designers can achieve a better speed up with a
smaller number of fast processors than a larger number of slower processors.
One important note that should be made that is closely related to Amdahl’s law is the fact that
some applications lack inherent parallelism, thus limiting the speedup that is achievable when
multiple processors are used. [6] [9] [14]
局部引用
Locality of Reference
The success of memory hierarchy is based upon assumptions that are critical to achieving
the appearance of a large, fast memory. The foundation of these assumptions is termed
locality of reference.
There are three components of the locality of reference, which coexist in an active process
which are:
时间
•
Temporal – A tendency for a process to reference in the near future the elements of
instructions or data referenced in the recent past. Program constructs that lead to
this concept are loops, temporary variables, or process stacks.
•
Spatial – A tendency for a process to reference instructions or data in a similar
location in the virtual address space of the last reference.
•
Sequentiality – The principle that if the last reference was rj(t), then there is a
likelihood that the next reference is to the immediate successor of the element rj(t).
空间
连续的
It should be noted that each process exhibits an individual characteristic with respect to the
three types of localities. [6] [18]
Cluster Computing
Page 35
Revision Version 2.3
3.1.4.
Copyright © 2003
Performance Analysis and Measurement
To understand complex interactions of many factors contributing to the performance of
parallel processors, performance measurement and analysis must be done. These are achieved
with specialised hardware and software. The traditional approach for measuring the
performance of a program running on a parallel computer is to time its execution from
beginning to end and compute the Megaflops (Mflops) and now more recently in Gigaflops
(Gflops) and Teraflops (Tflops). However timing alone is inadequate to fully characterize the
program’s behaviour. It is also necessary to detect and record data related to the dynamic
occurrences of hardware, system software, and application programming events and
interactions.
Measurements can be made on both hardware and software in both static and dynamic ways.
This information in this document concentrates on measuring Hardware and Software
performance in a dynamic way.
Some factors are discussed below that classify the types of performance measurement
available.
Hardware and Software Measurements
Hardware measurements focus on the physical events taking place within various machine
components. They are usually made with hardware monitors connected directly to hardware
devices, and are triggered by certain signals. Measuring devices need to be designed such that
they do not interfere with the systems being measured. For example when designing
microprocessors, it is important to make critical signals accessible for measurement. In this
instance chip manufacturers such as Intel have specific hardware registers that are able to
measure out-of-band performance. Additionally measurement instrumentation has been
integrated into the design of a few MPP systems.
Software Measurements are directed more towards logical events that occur during program
execution. These include individual timing of routines or tasks, and observations of events
related to loop-level parallelism, synchronisation, and other measurements. Special software is
needed to start, accumulate, and interpret records.
Static and Dynamic Analysis
The goal of static analysis is to extract, quantify, and analyse various characteristics of
benchmark code and it’s mapping to a particular architecture at compile time without actually
executing the code. Through static analysis we can measure the effectiveness of the compiled
portion of a system. In general, the purpose is to detect whether or not the software component
of the systems is responsible for poor performance. Conclusions may be reached about the
appropriateness of the hardware model, load algorithms, code restructuring, mapping and
allocation. The Paraphase compiler, developed at the University of Illinois, illustrates how
static analysis tools are used to perform tradeoffs.
The goal of dynamic analysis is to monitor the hardware and software while the application
program is running. Many performance losses are associated with inefficiencies in processor
scheduling, load balancing and data communication. [14]
3.1.5.
Practical Performance of Parallel Computers
While the mathematical models of section 3.1.3 serve to educate on the general performance of
parallelisation of sequential problems as well as inherently paralisable applications, it does not
take in to account Super-linear speedups and the high performance achievable with parallel
computers when dealing with search algorithms or highly parallel applications.
Beowulf systems are particularly well suited to, and may easily exploit process-level
parallelism that is exposed by running multiple, independent processes. [12]
Page 36
Cluster Computing
Revision Version 2.3
Copyright © 2003
When applicable in an application, process-level parallelism is often the easiest way to achieve
speedup in parallel computations. The only requirement for process-level parallelism is that
there must exist sequential code or codes that need to be run many times to produce a result.
This is otherwise known as parametric computing.
In general, the mathematical models discussed strictly dictate that to maximise speedup,
parallel systems should be built out of a small number of high performance processing
参数计算
elements. Whilst this is true for many applications, there are some notable examples where this
is not the case.
Eadline et al [38.1] describe their results with respect to I/O-dominant applications in a general
law that states:
“For two given parallel computers with the same cumulative CPU performance index, the
one which has slower processors has better performance for I/O-dominant applications”
The work shows that for I/O dominant applications it is better to have more data blocks being
processed at one time (by larger number of slower processors) than a smaller number of fast
processors.
Eadline’s work demonstrates that applications perform differently and performance cannot be
modelled effectively with mathematics at this stage, reinforcing the use of testing and
experimentation. Although this is the case, the popularisation of parallel computers will be
through the development of general-purpose, cost effective, fault-tolerant parallel clusters such
as Beowulf that can be tuned to a particular algorithm’s natural requirements or used in a
standard configuration.
Cluster Computing
Page 37
Revision Version 2.3
Copyright © 2003
3.2. Hardware Platforms
3.2.1.
CPU
A choice of CPU should be made from two families: Intel x86 [IA32] compatible (such as the
Pentium4) or Compaq Alpha systems [Formerly DEC]. Alternate vendor CPUs are supported
by Linux and cluster-able operating systems (such as the IBM Power PC chip and the SUN
SPARC), however there is only limited support and accompanying software distributions
available for these architectures. In general, Beowulf systems are mainly built with Intel or
Alpha architectures. Intel based systems are considered as commodity systems because there
are multiple sources (Intel, AMD, Cyrix) and obviously ubiquitous. Compaq Alpha on the
other hand, is a clear performance winner, but does represent a single source part, and
therefore can be hard to source at a good price.
Within the Intel based systems, the Pentium 3 and 4 [23] have the best floating point
performance and they support SMP motherboards. In addition, Intel has confirmed that the
Pentium 4 micro-architecture will scale to 10Ghertz over the next decade.
3.2.2.
Symmetric Multiprocessing
As previously discussed, Symmetric Multiprocessing (SMP) are tightly coupled
multiprocessors that communicate through a shared main memory. Hence, the rate at which
processors can communicate from one processor to the other is on the order of the bandwidth
of the memory. Depending on the hardware a small local cache may be used in each processor.
Symmetric Multiprocessor boards are commonly used in Beowulf clusters. In architectural
terms by adding SMP to cluster nodes we are using tightly coupled nodes in a loosely coupled
interconnection methodology.
The main advantages of using SMP in a cluster are:
1.
Lower price per performance.
2.
Greater communication speeds between processors on the same board (usually a
dual-processor configuration is used).
3.
Requires less space and draws less power.
The first advantage is important when building a large cluster. By using dual CPU SMP
systems across the whole cluster, only 50% of the network cards, cases, power supplies, and
motherboards are required. The only cost increase is a more expensive, SMP motherboard,
however the cost cuts far outweighs the increase.
If it is decided that SMP systems will not be used across the cluster, it is worthwhile to
consider an SMP server. The Topcat Research project [11] at the University of Technology,
Queensland has shown that it is quite common for a Beowulf master node to run at load of two
or higher, fully utilising both SMP CPUs. As the master node serves file systems to the client
nodes, it is important that the NFS server has enough CPU cycles left to perform its duties. A
workaround is to lower the nice level of nfsd (NFS Daemon), but to have spare CPU cycles is
also important. If during the operation of the cluster the server node will be loaded by users, it
should be considered to use a fast SMP system.
With SMP each node relies on the Linux (or the operating system of choice) internal
scheduler, which determines how the CPUs get shared. The user cannot (at this point) assign a
specific task to a specific SMP processor. The user can however, start two independent
processes or a threaded process on a particular SMP node and expect to see a performance
increase over a single CPU system.
Page 38
Cluster Computing
Revision Version 2.3
Copyright © 2003
Symmetric Multiprocessing
A typical SMP block diagram for an Intel system is shown below. Designed for scalability and
availability, 8-way Pentium III Xeon processor-based servers provide high performance and
price/performance for mid-range and back-end enterprise solutions – at a fraction of the cost of
proprietary RISC-based solutions [23.1]
Figure 3-4 – Intel SMP Server Block Diagram
Aside from the hardware architecture, the Software utilization of an SMP machine is through
the use of threads. Linux supports POSIX threads as do most other operating systems, but most
OS’s generally have their own enhanced thread packages.
Cluster Computing
Page 39
Revision Version 2.3
Copyright © 2003
3.2.3.
Basic Network Architectures
The design of a network system to suit a Beowulf cluster must take into account what type of
applications will be running on a cluster such as:
•
Degree of intercommunication required between each node during the running of an
application.
•
The effect of processor intercommunication on the throughput of the Cluster.
Once this is determined, it is a matter of deciding on the performance level required before
selecting an implementation topology and then technology.
Given the internal processor and memory structures of each node, distributed memory
architectures are primarily characterized by the network used to interconnect the nodes.
This network is usually represented as a graph, with:
•
Vertices corresponding to processor-memory nodes.
•
Edges corresponding to communications links.
If unidirectional communications links are used then directed-edges are used.
If bi-directional communications links are used then bi-directional edges are used. While bidirected edges signify bi-directional communication this represents both half-duplex and full
duplex communications.
Important parameters of an interconnection network include:
1.
Network Diameter – The longest of the shortest paths between various pairs of
nodes, which should be relatively small, is the latency is to be minimized. The
network diameter is more important with store-and-forward routing (when a message
is stored in its entirety and retransmitted by intermediate nodes) than with wormhole
routing (when a message is quickly relayed through a node in small pieces).
2.
Bisection (band)width – The smallest number (total capacity) of links that need to
be cut in order to divide the network into two sub networks of half the size. This is
important when nodes communicate with each other in a random fashion. A small
bisection (band)width limits the rate of data transfer between the two halves of the
network, thus affecting the performance of communications intensive algorithms.
3.
Vertex or node Degree – The number of communications ports required of each
node, which should be a constant, independent of network size if the architecture is to
be readily scalable to larger seizes. The node degree has a direct effect on the cost of
each node, with the effect being more significant for parallel ports containing several
wires or when the node is required to communicate over all ports at once.
网络直径
对分宽度
节点度
The following table lists these three parameters for some commonly used interconnection
networks. This table is by no means exhaustive but gives an idea of the variability of
parameters across different networks. [9]
Page 40
Cluster Computing
Revision Version 2.3
Copyright © 2003
No. of
Nodes
Network
Diameter
Bisection
Width
Node
Degree
Local
Links
1D Mesh (linear array)
k
k–1
1
2
Yes
1D Torus (ring, loop)
k
Network Topology
k/2
2
2
Yes
2
2k–2
k
4
Yes
2
k
2k
4
Yes1
3k-3
k2
k
2D Mesh
2D Torus (k-ary 2 cube)
k
3D Mesh
k3
3
3D Torus (k-ary 3 cube)
Pyramid
Butterfly
Yes1
9
No
3
No
2
2l
2 l −1
2 l −1
6
No
2l −1
l
2l
l
2k
(4k2–1)/3
2log2k
2k
1
2l–1
2l–2
2l (2 l +1 − 1)
2l
2 (l + 1)
2l
l
l
l
Cube-connected cycles
2l
Shuffle-exchange
2l
2l–1
De Bruijn
2l
l
1
6
3k/2
2
2l l
Hypercube
Yes
k
Binary Tree
4-ary hypertree
6
2
l +1
≥
4
No
l
No
3
No
4 unidir
No
4 unidir
No
With Folded Layout
Table 3-1 – Topological Parameters of Selected Interconnection Networks
While direct interconnection networks of the type shown in Table 3-1 have led to many
important classes of parallel systems, bus based architectures such as Ethernet dominate the
smaller scale parallel machines. However since a single bus can quickly become a
performance bottleneck as the number of processors increases, a variety of multiple-bus
architectures and hierarchical schemes are available for reducing bus traffic or increasing the
internode bandwidth.
Due to the ever-dropping prices of 100 Mbps Ethernet switches/NICs, fully meshed,
hypercube and the network topologies as outlined in Table 3-1 above are no longer economical
for Clustering.
Figure 2-7 depicts a typical Ethernet bus configuration interconnecting each Beowulf Cluster
node. 100 Mbps switched, full duplex Ethernet is the most commonly used network in
Beowulf systems, and gives almost the same performance as a fully meshed network. Unlike
standard Ethernet where all computers connected to the network compete for the cable and
cause collisions of packets, switched Ethernet provides dedicated bandwidth between any two
nodes connected to the switch.
As outlined in the following section should higher internode bandwidth be required, we can
use channel bonding to connect multiple channels of Ethernet to each node.
Cluster Computing
Page 41
Revision Version 2.3
Copyright © 2003
3.2.3.1.
Network Channel Bonding
Overview
One of the goals of a Beowulf system is to utilize scalable I/O using commodity subsystems.
For scaling network I/O a method was devised by NASA to join multiple low-cost networks
into a single logical network with higher bandwidth. The only additional work over a using
single network interface is the computationally simple task of distributing the packets over the
available device transmit queues.
Initially Beowulf clusters were designed to implement this method using two 10Mbps
Ethernets, later expanding it to three channels. Recent Beowulf systems have parallel 100Mpbs
Fast Ethernet channels or the latest Gigabit Ethernet channels.
Except for the performance increase, channel bonding is transparent to the applications
running on the cluster.
A by-product of the use of channel bonding is the redundancy of having two or more networks
interconnecting each node. Through the use of modified communications protocols this
network configuration can add fault tolerance and greater reliability to a cluster. [36]
A user-level control program called 'ifenslave' is available from NASA’s website [4]. Refer
Appendix B – Channel Bonding for the source listing and commentary.
The system-visible interface to "channel bonding" is the 'ifenslave' command. This command
is analogous to the 'ifconfig' command used to set up the primary network interface. The
'ifenslave' command copies the configuration of a "master" channel, commonly the lowest
number network, to slave channels. It can optionally configure the slave channels to run in a
receive-only mode, which is useful when initially configuring or shutting the down the
additional network interfaces.
Implementation
To minimize protocol overhead and allow the latest possible load-balancing decision, Beowulf
channel bonding is implemented at the device queue layer in the Linux kernel, below the IP
protocol level. This has several advantages:
•
Load-balancing may be done just before the hardware transmit queue, after logical
address (e.g. IP address) and physical address (e.g. Ethernet station address) are
added to the frame.
•
Fragmented packets, usually large UDP/IP packets, take advantage of the multiple
paths.
•
Networks that fail completely first fill their hardware queues, and are subsequently
ignored.
This implementation results in the following limitations:
•
All connected machines must be on the same set of bonded networks. The whole
cluster must be channel bonded as communication between a channel bonded node
and a non-channel bonded node is very slow if not impossible.
•
It's difficult to detect and compensate if one network segments (splits in the middle).
•
Channel bonding shutdown is not as simple. The safest way to restore single channel
performance is to either reboot each system or use the network manager (part of the
control panel) to shutdown and restart each interface.
The software implementation is divided into two parts, the kernel changes needed to support
channel bonding, and the user-level control program 'ifenslave'.
•
Page 42
The user-level control program 'ifenslave' is compiled from the source ifenslave.c.
Cluster Computing
Revision Version 2.3
Copyright © 2003
•
Apply the kernel patch (linux-2.0.36-channel-bonding.patch from Paralogic [38]),
run xconfig and enable Beowulf Channel bonding then rebuild and install the kernel.
Typical requirements of the hardware configuration is shown as follows:
•
n Ethernet NICs per system.
•
n hubs (one for each channel) OR n switches (one for each channel) OR a switch that
can be segmented into n virtual LANs.
Once a cluster is channel bonded it can be tested by running the network performance
benchmark netperf or similar. [4.1] [38]
3.2.4.
Node Interconnection Technologies
A node interconnection technology is the hardware and software interface that connects each
node and allows node intercommunication, as distinct from the interconnect topology as
discussed in 3.2.3. This document concentrates on Class 1 Beowulf clusters, which has
implications for the node interconnect used [Refer to 2.3.3 and Appendix 8.1 for details].
Some design considerations for the selection of a node interconnect are as follows:
•
Linux support: yes/no, kernel driver or library (kernel drivers are preferred).
•
Maximum bandwidth: The higher the better.
•
Minimum latency: The lower the better.
•
Available as: Single Vendor / Multiple Vendor Hardware.
•
Interface port/bus used: High performance, included as a standard node port and
matched bandwidth to the dedicate node network fabric bandwidth.
•
Network structure: Bus/Switched/Topology.
•
Cost per machine connected: The lower the better.
A full analysis of the available node interconnection technologies are presented in Appendix
8.1.
As previously stated 100 Mbps switched, full duplex Ethernet is the most commonly used
network in Beowulf systems as it:
1.
Gives almost the same performance as a fully meshed network.
2.
Is one of the most commonly available networks.
3.
Cost effective.
As such it will be used for all testing and experimentation in this document. Refer to
Appendix 8.1 for technical details.
Cluster Computing
Page 43
Revision Version 2.3
Copyright © 2003
3.3. Operating Systems
3.3.1.
General
A modern operating system provides two fundamental services for users:
1.
Provides a Virtual Machine – A programming and user interface that hides the
obscurities of using/programming the various hardware components that make up a
system.
2.
Shares hardware resources among users or executing processes – One of the most
important resources is the processor. In all multitasking operating systems (which are
the only type of operating system dealt with in this document) the computer performs
multiple tasks at once by scheduling jobs using various scheduling algorithms which
allows many processes and threads to run concurrently in pseudo parallel (as with the
case of a one processor machine) or parallel (as with the case with an SMP machine).
隐匿
These services are provided to the user by the operating system through the following
operating system functions:
3.3.2.
1.
Process Management – The operating system must allocate resources to processes,
enable processes to share and exchange data, protect the resources of each process
from one another, enable synchronisation among processes, and start, stop and
suspend processes according to different priorities and scheduling algorithms.
2.
Memory Management – The operating system needs to handle the relocation,
protection, sharing, logical and physical organization of memory through various
schemes.
3.
File Systems Management – The primary consideration for an operating system is
the need to manage local files including structure, organization, naming, access,
allocation and locking.
Distributed systems as detailed in this document are a special case where the
operating system and/or an additional layer of middleware is required to manage file
systems across a network, where files can be distributed across disjointed memory
spaces.
4.
I/O – An operating system needs to interface with many different forms of hardware
to collect data and communicate results. This requires integration, logical
structuring, control, intercommunication and programming abilities. I/O can be
roughly grouped into three subcategories:
1.
Human Readable – i.e. Video display terminals, printers, keyboard, mouse
and speakers.
2.
Machine Readable – i.e. disk drives, digital camera and other devices.
3.
Communication – i.e. Modems & network cards.
Towards Parallel Systems
The operating system is a critical factor to the development of parallel systems. In all Beowulf
systems, a required software layer called middleware augments the operating system.
Middleware handles source code programming, compilation, and parallel execution and
runtime requirements.
However, the latest generation of operating systems are emerging with support for parallel
systems paradigms by integrating middleware functions into the operating system. Three of the
most commonly used operating systems (Microsoft Windows, SUN Solaris and Redhat Linux)
are including kernel level support for parallel programming, inherent parallelisation and load
Page 44
Cluster Computing
Revision Version 2.3
Copyright © 2003
balancing characteristics. This is the next leap forward in parallel systems as it caters for the
corporate environment and will eventually pervade into desktop machines.
The main issues that these so-called next generation operating systems need to address (and
currently do so to varying degrees) to exploit the cluster environment further than scientific
computing capabilities, are listed below: [32]
•
Failure Management – These include fault tolerant or highly available solutions.
Fault tolerant solutions will recover transparently to an application any internal node
or component failure. In contrast, highly available solutions offer a high probability
that all nodes will be in service. However, with highly available solutions any
transactions lost in the process of a node or component failure need to be handled or
recovered at the applications layer (such as with atomic transactions).
•
Load Balancing – Clusters require an effective capability for balancing the load
among available computers. This includes the ability for the system to cater in terms
of dynamic task scheduling and reconfiguration for the addition or removal of nodes
from the cluster.
•
Parallelising Computation – In some cases, as with the case with scientific
computing, effective use of a cluster requires executing software from a single
application in parallel. There are three general approaches to this problem:
故障管理
负载平衡
并行计算
1.
Parallelizing Compiler – A parallel compiler determines at compile time,
which parts of an application can be executed in parallel. These are then split
off and assigned to different computers in the cluster. Performance depends
on the nature of the problem and how well the compiler is designed.
2.
Parallel and Cluster Aware Applications – In this approach the
programmer write an application from the outset to run on a cluster, and uses
message passing to move data, as required, between nodes. This places a
high burden on the programmer but may be the best approach for exploiting
clusters for some applications. Another technique that can be used is to write
a multi-threaded cluster aware application and let the cluster manage the
threads.
并行编译
适宜于并行和机群的应用
3.
参数计算
Parametric Computing – This approach can be used if the problem is
inherently parallel, where an algorithm or program is executed a large
number of times, each time with a different set of starting conditions or
parameters. A good example is simulation programs, however tools are
required to organise, run and manage jobs in an orderly manner.
These next generation Operating Systems use one of the following two cluster implementation
models, currently used in the computer industry – the shared device and shared nothing models
which are described below:
Shared Device Model
In the shared device model, software running on any computer in the cluster can gain access to
any hardware resource connected to any computer in the cluster (for example, a hard drive). If
two applications require access to the same data, much like a symmetric multiprocessor (SMP)
computer, access to the data must be synchronized. In most shared device cluster
implementations, a component called a Distributed Lock Manager (DLM) is used to handle
this synchronization.
The DLM is a service that is provided to applications running on the cluster that keeps track of
references to hardware resources within the cluster. If multiple applications attempt to
reference a single hardware resource, the DLM recognizes and resolves the conflict. However,
using a DLM introduces a certain amount of overhead into the system in the form of additional
message traffic between nodes of the cluster as well as the performance loss due to serialized
access to hardware resources.
Cluster Computing
Page 45
Revision Version 2.3
Copyright © 2003
Shared Nothing Model
The shared nothing model is designed to avoid the overhead used by the DLM in the shared
device model. In this model, each node of the cluster owns a subset of the hardware resources
that make up the cluster. As a result, only one node can own and access a hardware resource at
a time. When a failure occurs, another node takes ownership of the hardware resource so that
the hardware resource can still be accessed.
In this model, requests from applications are automatically routed to the system that owns the
resource. For example, if an application needs to access a database on a hard drive owned by a
node other than the one it is running on, the node passes the request for the data to the other
node. This allows the creation of applications that are distributed across nodes in a cluster.
3.3.3.
Implementations
The rest of this section introduces three operating system implementations that constitute the
majority of clusters currently being implemented:
1.
Linux Redhat version 7.2 based on the 2.4.x series kernel.
2.
Microsoft Windows 2000.
3.
SUN Solaris.
Within each section the standard version of the OS is described, as well as the clustering
version. It is of note that in all cases the clustering version costs more than the standard version
and hence, all scientific computing clusters are built using the standard versions (to maximise
computing power for cost) and corporations tend towards the cluster versions (as they require
the inbuilt clustering functions, reduced administration overheads, support and cost is not
such a significant issue).
The last heading in the section refers briefly to other operating systems that have been used in
clustering.
3.3.4.
Redhat Linux 7.2
Redhat Linux is a distribution of Linux that has received over a billion dollars worth of capital
from companies such as IBM to produce an operating system to rival the Windows platform. 与。。。对抗
Whilst there are many distributions of the Linux Operating system, Redhat Linux is the most
popular according to International Data Corp. (IDC) research [22]. The Linux Redhat
distribution also introduced the Redhat Package Manager and according file packages denoted
by the .rpm extension. This system allows precise control of installing and removing
packages (programs utilities, etc).
As each different Linux distribution uses its own numbering schema is can be difficult to know
which Linux version is actually implemented. Hence, the crucial number to check is the Linux
Kernel version.
The latest version available is Redhat Linux 8.0 [Kernel 2.4.18]
For laboratory testing purposes the Linux Redhat 7.2 distribution will be used, implementing
the 2.4.7-10 kernel. For further implementation details of the Linux Operating system refer to
4.1 for further details.
Features of the Linux Operating Systems include:
Page 46
•
Every line of the source code for the kernel and utilities is available at no cost.
•
There are no licensing fees to run Linux. Refer to the GNU General Public License
[39].
•
Demand Paging.
•
Virtual Memory.
Cluster Computing
Revision Version 2.3
Copyright © 2003
•
Multitasking.
•
Multiple File Systems Support.
•
Networking Support.
•
Program Development Tools.
•
Dynamically loadable, shared libraries.
•
Graphical User Interface.
•
Robust
Refer to Figure 3-5 for a basic block diagram of the structure of Linux.
Figure 3-5 – System Component Relationships in the Linux Operating System
The Linux kernel interacts with file systems through a layer called the Virtual File System
(VFS). This allows Linux to easily incorporate new file systems as the VFS provides the
operating system a standard interface layer which mounts, reads, and writes file systems
without requiring individual specific file formats.
Linux uses the Extended-2 File system, EXT-2, however with the release of Linux Redhat 7.2
and the 2.4.7-10 Kernel, the EXT3 file system is available.
File systems are fundamentally tree structures, but EXT2, like most Unix file systems, allows
the file system to form directed graphs with hard linking and symbolic links as shown in
Figure 3-6. [12] [31]
Cluster Computing
Page 47
Revision Version 2.3
Copyright © 2003
Figure 3-6 – Linux Hierarchical File System Structure
With respect to parallel systems, currently a parallel file system is available for Linux (at the
time of writing only available in beta format) known as Parallel Virtual File System (PVFS).
The Parallel Virtual File System (PVFS) project is an effort to provide a high-performance and
scalable parallel file system for PC clusters. PVFS is open source and released under the GNU
General Public License [39]. It requires no special hardware or modifications to the kernel.
PVFS provides four important capabilities in one package: [40]
•
A consistent file name space across the machine.
•
Transparent access for existing utilities.
•
Physical distribution of data across multiple disks in multiple cluster nodes.
•
High-performance user space access for applications.
Redhat High Availability Server
The investment of IBM in Redhat Linux, has allowed the vendor to develop a Clustering
solution to rival that of Windows Cluster Server to deliver advanced security, Scalability,
availability and reliability.
Known as Red HatTM High Availability Server, it is a specialized version of the commonly
used Red Hat Linux distribution, however, with a price of USD $1995. [22]
Red Hat High Availability Server is an out-of-the-box clustering solution that delivers
dynamic load balancing, improved fault tolerance and scalability of TCP/IP based applications.
It lets users combine individual servers into a cluster, resulting in highly available access to
critical network resources such as data, applications, network services, and more. If one server
in the cluster fails, another will automatically take over its workload. The Red Hat High
Availability Server is ideally suited to Web servers, ftp servers, mail gateways, firewalls, VPN
gateways and other front-end IP-based applications where virtually uninterrupted service is
required.
The product supports heterogeneous network environments, allowing individual members of
the cluster to run Red Hat Linux or virtually any other OS including Solaris®, and Windows
NT®. Because the Red Hat High Availability Server is an open source product, customers are
free from expensive technology lock-in that often occurs with proprietary solutions.
Page 48
Cluster Computing
Revision Version 2.3
Copyright © 2003
Security Features
The Red Hat High Availability Server has a number of inherent security features designed
specifically for high availability Web front-end applications. Remote system access is disabled
by default, and unused network services are not installed or started in the standard installation.
The Red Hat High Availability Server can be configured in two main ways. In Failover
Services (FOS) mode, the system can be configured as a two node cold failover cluster ideally
suited for applications where simple, affordable redundancy is needed such as firewalls, static
Web servers, DNS, and mail servers. In Linux Virtual Server (LVS) mode, the system can be
configured as an n-node cluster consisting of a two node load balancer, which accepts requests
and directs those request to one of any number of IP-based servers based on a configurable
traffic management algorithm.
Red Hat High Availability Server Features and Benefits
•
Ease of Installation – the installation manager installs only those packages that are
needed with the clustering packages.
•
High Performance and Scalability – The Red Hat High Availability Server supports the
scalability that meets the growth demands of highly dynamic IP environments. The
number of cluster nodes is limited only by the hardware and network used. The product
has advanced cluster features that provide high levels of performance including an ability
to configure servers to bypass the load balancers when returning traffic back to the client,
increasing the overall performance of the cluster. Additionally, because individual nodes
can be taken off-line for maintenance or upgrades without interruption of service costly
downtime can be eliminated.
•
Maximized Flexibility – The Red Hat High Availability Server offers System
Administrators a high degree of flexibility. In LVS mode the product supports four load
balancing methods (Round Robin, Weighted Round Robin, Least Connections, and
Weighted Least Connections) and three traffic forwarding techniques (IP Masquerading,
Tunnelling and Direct Routing). Virtually every popular IP service is supported including
Web (http), email, FTP, LDAP, DNS, as well as others.
•
Increased Security – The Red Hat High Availability Server has built-in security features
designed to withstand common attacks. System Administrators can set-up sand traps,
providing for redirection of IP traffic from a potential attacker to a secure address. Out of
the box, finger, talk, wall, and other daemons are disabled or not installed. In addition,
multiple traffic routing and scheduling techniques along with virtual IP addresses allow
creation of a security barrier for the network.
•
Availability – The Red Hat High Availability Server dramatically reduces the likelihood
of system failure by quickly detecting component server and application failures and
redirecting workload to the remaining servers in the cluster. If one or more servers fail,
others take over with minimal interruption.
•
Cost Effective – Uses inexpensive commodity hardware to lower your overall cost of
purchasing and maintaining the system.
•
Support – A one-year support package that includes standard hours installation and
configuration assistance and 24x7 server-down support.
Cluster Computing
Page 49
Revision Version 2.3
Copyright © 2003
3.3.5.
Microsoft Windows 2000
Microsoft Windows is a dominant operating system in the PC marketplace. It is based on the
Windows NT architecture and it is a stable, multitasking, priority-based pre-emptive
scheduling, and multi-threaded 32-bit operating system.
Some of the features of W2K are listed below:
•
Supports multiple CPU’s and provides multi-tasking using SMP.
•
Supports different CPU’s using the HAL (NT previously supported Intel x86, DEC
Alpha, MIPS and it is understood this is set to continue) and multiprocessor machines
using POSIX threads.
•
Uses an object-based security model and its own file system (NTFS) that allows
permissions to be set on a file and directory basis.
•
Built-in Networking protocols such as IPX/SPX, TCP/IP, and NetBEUI.
Figure 3-7 below illustrates the overall structure of W2K.
Figure 3-7 – Windows 2000 Architecture
As with the majority of operating systems, W2K separates application-oriented software from
operating systems software. The latter, which includes the Executive, the micro-kernel device
drivers, and the HAL, runs in kernel mode. Kernel mode software has access to system data
and to the hardware. The remaining software, running in user mode, has limited access to
system data.
W2K has what is referred to as a modified micro-kernel architecture, which is highly modular.
Each system function is managed by just one component of the operating system and all
applications access that function through the responsible component using a standard interface.
In principle any module can be removed, upgraded, or replaced without rewriting the OS or
any of its standard API’s.
However, unlike a pure micro-kernel system, W2K is configured so that many of the system
functions out-side the micro-kernel run in kernel mode. The reason is to boost the
performance.
推进
Page 50
Cluster Computing
Revision Version 2.3
Copyright © 2003
The most important sub-system is the Win32 API. The Win32 API allows the applications
programmer access to the many functions of the Windows platform and is the point of
interface for the cluster middleware layer of software. [31] [32]
One of the drawbacks of using the Windows 2000 platform (standard and clustering versions)
for clustering is cost. A cluster is designed to function as a single machine, however in most
cases we use a local operating system on each node of the system and with the W2K licensing
model this would mean we would need to pay a licence cost per node. One possible way
around this is to use diskless nodes. However, with operating systems such as Linux that are
freely available under the GNU General Public License model [39] this is not required.
Windows 2000 Cluster Server
Windows 2000 (W2K) Cluster Server (formerly codenamed Wolfpack) is a shared-nothing
cluster, in which each disk volume and other resources are owned by a single system at a time.
A typical set-up for a W2K cluster is shown in Figure 3-8 below.
Figure 3-8 – Windows 2000 Cluster Server Topological Diagram
W2K is the next generation from Windows NT 4.0 and is the second generation of cluster
operating systems from Microsoft.
W2K Cluster Server is based on the following concepts: [32]
•
Cluster Service – The collection of software on each node that manages all clusterspecific activity.
•
Resource – An item managed by the cluster service. All resources are objects
representing actual resources in the system, including physical hardware devices such
as disk drives, network cards and logical items such as logical disk volumes, TCP/IP
addresses, entire applications, and databases.
•
Online – A resource is said to be online at a node when it is providing service on that
specific node.
•
Group – A collection of resources managed as a single unit. Usually, a group
contains all of the elements needed to run a specific application and for client systems
to connect to the service provided by that application.
The Cluster Service (Clussvc.exe) and Resource Monitors (Resrcmon.exe) are the keys to a
Cluster Server cluster. Each node in a cluster contains all of the components shown in Figure
3-9 and described in the architectural overview below. [34]
Cluster Computing
Page 51
Revision Version 2.3
Copyright © 2003
Architectural Overview
Figure 3-9 – Windows 2000 Cluster Server Block Diagram
The Database Manager is the component of the Cluster Service that implements the cluster
database. This database contains information about all of the entities in the cluster, such as the
cluster itself, resource types, groups, and resources. The cluster database is stored in the
registry on each node of the cluster.
The Database Managers, one on each node in the cluster, cooperate with each other to maintain
consistent cluster configuration information.
In addition, the Database Manager also provides an interface to the configuration database that
is used by the other components of the Cluster Service. Cluster services coordinate updates to
the registry to maintain consistency and provide atomic updates across the nodes in the cluster.
The Node Manager on one node in the cluster communicates with the Node Manager on the
other node in order to detect node failures in the cluster. This is accomplished by sending
"heartbeat" messages between the nodes.
If the Node Manager on a node does not respond, the active resources must be failed over to
the node that is still running. The Resource/Failover Manager does the failover portion of this
process. If the Cluster Service fails on a node, all of the resources on the node will failover
even if the resources are still all online and functioning. This happens as a result of the Node
Manager initiating failover of the resources since it is unable to communicate with the Node
Manager on the other node.
法定人数
选定的团体
Page 52
Quorum Resource Interaction – If the communication link between the nodes becomes
unavailable but both nodes are still functioning, the Node Manager on each node would try to
fail the resources over to the node on which the Node Manager was running. This would result
in each node thinking it is the surviving node of the cluster and would try to bring all the
resources online. To prevent this situation, the Cluster Server relies on the quorum resource to
ensure that only one node has a resource online. If communication between nodes fails, the
node that has control of the quorum resource will bring all resources online. The node that
Cluster Computing
Revision Version 2.3
Copyright © 2003
cannot access the quorum resource and is out of communication will take any resources it had
offline.
The Event Processor is the communications centre of the Cluster Service. It is responsible for
connecting events to applications and other components of the Cluster Service. This includes:
•
Maintenance of cluster objects.
•
Application requests to open, close, or enumerate cluster objects.
•
Delivery of signal events to cluster-aware applications and other components of the
Cluster Service.
The Event Processor is also responsible for starting the Cluster Service and bringing the node
to the "offline" state. The Event Processor then calls the Node Manager to begin the process of
joining or forming a cluster.
Communication Manager – The components of the Cluster Service communicate with the
Cluster Service on other nodes through the Communication Manager. Communication
Manager is responsible for the following:
•
KeepAlive protocol. Checks for failure of the Cluster Service on other nodes.
•
Resource group push. Initiates failover of resources from one node to another node.
•
Resource ownership negotiation. Determines the new owner of resources from a
failed node.
•
Resource state transitions. Nodes notify the rest of the cluster when a resource goes
offline or comes online. This is used to help track the owner of a resource.
•
Enlist in a cluster. Initial contact with a cluster.
•
Join a cluster. Synchronization with the other node in the cluster after going offline
and coming back online.
•
Database updates. Two-phase commit of cluster database updates.
The Global Update Manager provides an interface for other components of the Cluster
Service to initiate and manage updates. The Global Update Manager allows for changes in the
online/offline state of resources to be easily propagated throughout the nodes in a cluster. In
addition, notifications of cluster state changes are also sent to all active nodes in the cluster.
Centralizing the global update code in a single component allows the other components of the
Cluster Service to use a single, reliable mechanism.
The Resource/Failover Manager is responsible for:
•
Managing resource dependencies.
•
Starting and stopping resources, by directing the Resource Monitors to bring
resources online and offline.
•
Initiating failover and failback.
In order to perform the preceding tasks, the Resource/Failover Manager receives resource and
cluster state information from the Resource Monitors and the Node Manager. If a resource
becomes unavailable, the Resource/Failover Manager either attempts to restart the resource on
the same node or initiates a failover of the resource.
Cluster Computing
Page 53
Revision Version 2.3
Copyright © 2003
3.3.6.
Sun Solaris
The Solaris operating system from SunSoft is a UNIX-based multithreaded and multi-user
operating system. It supports Intel x86 and SPARC-based platforms. Its network support
includes a TCP/IP protocol stack and layered features such as Remote Procedure Calls (RPC),
and the Network File System (NFS). The Solaris programming environment includes ANSIcompliant C and C++ compilers, as well as tools to profile and debug multithreaded programs.
The Solaris kernel supports multithreading, multiprocessing and has real-time scheduling
features that are critical for multimedia applications. Solaris supports two kinds of threads:
Light Weight Processes (LWPs) and user level threads (ULTs). [31]
In addition to many other features, Solaris has an in built Clustering feature set, which is of
interest in both parallel processing and for reliability purposes.
Sun Cluster
Sun Cluster is a distributed operating system built as a set of extensions to the base Solaris
UNIX system. It provides the cluster with a single-system image; that is the cluster appears to
the user and applications as a single computer running the Solaris operating system.
Figure 3-10 shows the overall architecture of Sun Cluster. The major components are as
follows: [32]
•
Object and communication support – The SUN Cluster implementation is object
oriented. The CORBA object model is used to define objects and the remote
procedure call mechanism.
•
Process management – Global process management extends process operations so
that the location of a process is transparent to the user. Process migration is possible,
i.e. a process can move from one node to another during its lifetime to achieve load
balancing or for failover.
•
Networking – SUN Cluster implements networking systems to support the cluster
implementation.
•
Global distributed file system – SUN Cluster implements a global file system to
simplify file system operations and process management over the entire cluster.
Figure 3-10 – Sun Cluster Structure
3.3.7.
Other
Whilst Linux, Windows NT and SUN Solaris have been discussed with respect to clustering,
this is not to say that other Operating systems are not available that can be used in a cluster
configuration.
For reference, some of the other operating systems that have been used in this configuration
are listed below:
Page 54
Cluster Computing
Revision Version 2.3
Copyright © 2003
•
Free BSD – Based on the Berkley Standard Distribution of Unix produced by the
University of California at Berkley. This operating system is freely available.
•
MAC OS – The Macintosh Operating system for use with Macintosh Hardware.
•
Scyld Beowulf Cluster Operating System – The original Beowulf creators and
researchers [2.3] formed Scyld Corporation and are producing pre-configured cluster
suited Linux distributions. As can be expected these Linux distributions cost as much
as $8,000 US dollars for a 32-processor cluster as opposed to plain Linux available
free under the GPL [27].
3.4. Middleware
Middleware is the key differentiator between a standard network PC’s and a parallel
processing supercomputing facility. Middleware is a software layer that is added on top of the
operating system to provide what is known as a Single System Image (SSI).
Middleware provides a layer of software that enables uniform access to different nodes on a
cluster regardless of the operating system running on a particular node. This is a similar
concept to the Hardware Abstraction Layer (HAL) in Windows 2000 that provides uniform
and portable access to a PC’s hardware.
The middleware is responsible for providing high availability, by means of load balancing and
responding to failures in individual components.
The following lists the desirable objectives of cluster middleware services and functions [32]:
1.
Single Entry Point – A user logs onto the cluster rather than on to an individual
computer.
2.
Single File Hierarchy – The user sees a single hierarchy of file directories under
the same root directory.
3.
Single Control Point – There is a default workstation used for cluster
management and control. This is usually known as the server.
4.
Single Virtual Networking – Any node can access any other point in the cluster,
even though the actual cluster configuration may consist of multiple
interconnected networks. There is a single virtual network operation.
5.
Single Memory Space – Distributed Shared Memory enables programs to share
variables.
6.
Single job-management system – Under a cluster job scheduler, a user can
submit a job without specifying the host computer to execute the job.
7.
Single User Interface – A common graphic interface supports all users,
regardless of the workstation from which they enter the cluster.
8.
Single I/O Space – Any node can remotely access any I/O peripheral or disk
device without knowledge of its physical location.
9.
Single Process Space – A uniform process-identification scheme is used. A
process on any node can create or communicate with any other process on a
remote node.
10. Checkpointing – This function periodically saves the process state and
intermediate computing results, to allow rollback recovery after a failure.
11. Process Migration – This function enables load balancing.
Items 1 through 7 are concerned with providing a Single System Image, and items 8 through
11 are concerned with enhancing the availability of the cluster.
This section details the recent advances in Cluster Middleware.
Cluster Computing
Page 55
Revision Version 2.3
3.4.1.
Copyright © 2003
Parallel Communications Libraries
A cluster is a collection of local memory machines. The only way for Node A to
communication to Node B is through the network. Software built on top of this architecture
‘passes messages’ between nodes. While message-passing codes are conceptually simple, their
operation and debugging can be quite complex.
A major goal of middleware for message passing systems is to ensure the degree of portability
across different machines. The expectation is that the same message passing code should be
able to be executed on a variety of machines as long as the message-passing library is
available. Whilst this is the case, some tuning may be required to take best advantage of the
features of each system.
There are two popular message passing libraries (or lowest layer of middleware) that are
commonly used are PVM and MPI. Both PVM and MPI provide a portable software API that
supports message passing.
From a historical stand point PVM appeared first and was designed to work on networks of
workstations (to create a Parallel Virtual Machine). It has since been adapted to many parallel
supercomputers (using both distributed and shared memory). Control of PVM is primarily
with its authors.
In contrast, MPI is a standard that is supported by many hardware vendors (such as IBM, HP,
HITACHI, SUN, & Cray). It provides more functionality than PVM and has versions for
networks of workstations (such as Beowulf clusters). Control of MPI is with the standards
committee known as the MPI forum. The most current release of the MPI standard is MPI2.
[3] [11]
As MPI2 is the current standard for message passing systems, the experimental testing phase
of this documentation focuses on this library although both PVM and MPI are discussed.
3.4.1.1.
PVM Overview
The PVM (Parallel Virtual Machine) software provides a unified framework within which
parallel programs can be developed in an efficient and straightforward manner using existing
hardware. PVM enables a collection of heterogeneous computer systems to be viewed as a
single parallel virtual machine. PVM transparently handles all message routing, data
conversion, and task scheduling across a network of incompatible computer architectures.
The PVM computing model is simple yet very general, and accommodates a wide variety of
application program structures. The programming interface is deliberately straight forward,
thus permitting simple program structures to be implemented in an intuitive manner. The user
writes his application as a collection of cooperating tasks. Tasks access PVM resources
through a library of standard interface routines. These routines allow the initiation and
termination of tasks across the network as well as communication and synchronization
between tasks. The PVM message-passing primitives are oriented towards heterogeneous
operation, involving strongly typed constructs for buffering and transmission. Communication
constructs include those for sending and receiving data structures as well as high-level
primitives such as broadcast, barrier synchronization, and global sum.
PVM tasks may possess arbitrary control and dependency structures. In other words, at any
point in the execution of a concurrent application, any task in existence may start or stop other
tasks or add or delete computers from the virtual machine. Any process may communicate
and/or synchronize with any other. Any specific control and dependency structure may be
implemented under the PVM system by appropriate use of PVM constructs and host language
flow-control statements. Owing to its ubiquitous nature (specifically, the virtual machine
concept) and because of its simple but complete programming interface, the PVM system has
gained widespread acceptance in the high-performance scientific computing community. [2]
PVM contains fault tolerant features that become more important as the cluster size grows.
The ability to write long running PVM applications that can continue even when hosts or tasks
Page 56
Cluster Computing
Revision Version 2.3
Copyright © 2003
fail, or loads change dynamically due to outside influence, is quite important to heterogeneous
distributed computing. [21]
PVM documentation and distributions for both Windows NT and Unix/Linux are available
from the PVM home page at: http://www.epm.ornl.gov/pvm/.
The most current version of PVM is 3.4.3.
It is also noted that due to PVM’s long running history there are many other websites that
contain PVM source code and programming examples.
3.4.1.2.
MPI Overview
MPI (Message Passing Interface) is a standardized and portable message-passing system
designed by a group of researchers from academia and industry to function on a wide variety
of parallel computers. The standard defines the syntax and semantics of a core of library
routines useful to a wide range of users writing portable message-passing programs in Fortran
77, C and now C++ and Fortran 90. Several well-tested and efficient implementations of MPI
already exist, including some that are free and in the public domain. These are beginning to
foster the development of a parallel software industry, and there is excitement among
computing researchers and vendors that the development of portable and scalable, large-scale
parallel applications is now feasible.
10 Reasons for use of MPI over PVM
Many comparisons of the two message passing models have been undertaken in the Research
Community [19.1], [21] and the salient reasons for using MPI are outlined below: 突出的
1.
MPI has more than one freely available, quality implementation – There are at least
LAM, MPICH and CHIMP. The choice of development tools is not coupled to the
programming interface.
2.
MPI defines a 3rd party profiling mechanism – A tool builder can extract profile
information from MPI applications by supplying the MPI standard profile interface in a
separate library, without ever having access to the source code of the main
implementation.
3.
MPI has full asynchronous communication – Immediate send and receive operations
can fully overlap computation.
4.
MPI groups are solid, efficient, and deterministic – Group membership is static. There
are no race conditions caused by processes independently entering and leaving a group.
New group formation is collective and group membership information is distributed, not
centralized.
5.
MPI efficiently manages message buffers – Messages are sent and received from user
data structures, not from staging buffers within the communication library. In some cases
Buffering may be totally avoided.
6.
MPI synchronization protects the user from third party software – All
communication within a particular group of processes is marked with an extra
synchronization variable, allocated by the system. Independent software products within
the same process do not have to worry about allocating message tags.
7.
MPI can efficiently program MPP and clusters – A virtual topology reflecting the
communication pattern of the application can be associated with a group of processes. An
MPP implementation of MPI could use that information to match processes to processors
in a way that optimises communication paths.
多种开发工具
第三方工具
异步通讯
组坚实、有效、确定性
有效管理消息缓冲
同步保护第三方软件
可有效应用于MPP和机群
8.
MPI is highly portable – Recompile and run on any implementation. With virtual
topologies and efficient buffer management, for example, an application moving from a
cluster to an MPP could even expect good performance.
Cluster Computing
Page 57
高度轻便
Revision Version 2.3
形式化说明
标准
9.
Copyright © 2003
MPI is formally specified – Implementations have to live up to a published document of
precise semantics.
10. MPI is a standard – Its features and behaviour were arrived at by consensus in an open
forum. It can change only by the same process.
The most current MPI standard, MPI2 can be obtained from the MPI forum website [20] at:
http://www.mpi-forum.org/
The most notable additions to the MPI2 standard (from the original MPI standard) are listed
below:
•
Parallel I/O
•
Remote Memory Operations
•
Dynamic Process Management
•
Supports the use of threads – It is noted that POSIX Threads (as defined by the
POSIX standard) should be used to ensure portability.
While the MPI standards are freely available from the MPI forum, distributions for different
operating systems and platforms are available from various groups. A reliable source that lists
all current commercial and freely available distributions is the LAM-MPI home page at:
http://www.lam-mpi.org/mpi/implementations/
Two of freely available distributions of MPI are known as MPICH and LAM. Many
performance comparisons of the two have been undertaken in the Research Community [19.4].
For testing purposes the LAM distribution will be used as in many cases LAM exhibits better
performance [19.4] and has greater compliance with the MPI 2 standard [19.2]. Refer to
Appendix C for Standard compliance details and the LAM-MPI website [19.4] for
performance details.
LAM – Local Area Multicomputer
LAM (Local Area Multicomputer) is an MPI programming environment and development
system for heterogeneous computers on a network. With LAM, a dedicated cluster or an
existing network computing infrastructure can act as one parallel computer solving one
problem.
The latest stable version of LAM/MPI is 6.5.9
[Released 28 Jan 2003]
LAM features extensive debugging support in the application development cycle and peak
performance for production applications. LAM features a full implementation of the MPI
communication standard.
There are currently two versions of LAM available: LAM 6.5.9 and LAM 6.6 beta:
•
LAM version 6.5.9 includes many new features (such as support for Interoperable
MPI [IMPI]) and bug fixes. It is the most recent stable release. More information on
IMPI can be found in [3].
•
LAM version 6.6 beta is the next generation LAM software, except it is only available
in beta release and hence is not able to be used for a production cluster.
LAM on Linux
As reported by the MPI-LAM Development Team [19.3] the 2.2.x series of the Linux kernel
has bugs in part of its TCP/IP implementation. This has caused performance problems in the
LAM/MPI software. Linux 2.2.10 appears to have rectified this problem, since LAM
performance in 2.2.10 is comparable to LAM performance under 2.0.36. The LAM
development team have not found a way to explain the performance differences noted in the
TCP test program, even in 2.2.10.
Page 58
Cluster Computing
Revision Version 2.3
Copyright © 2003
All Linux versions that shall be used will be Redhat Linux Version 6.2, kernel version 2.2.145.0 and later. [22]
3.4.2.
障碍
Application Development Packages
On of the major roadblocks to parallel processing are the billions of dollars worth of existing
software that are only able to be run on serial SISD machines.
The cost of the change to switch to parallel programs and retraining of programmers is
prohibitive. Whilst this objection is true and it is noted that not all programs needed in the
future have been written. New applications will be developed and many new problems will
become solvable with increased performance through the use of parallel systems. Students of
operating systems are already being trained to think in parallel and use methodologies that can
parallelise serial applications. Additionally, tools are both currently available and being
developed to transform sequential code into parallel code automatically (Fortran and C tools
are available [11]).
In a pre-emptive initiative, it has been suggested by Behrooz [9] that it may be prudent to 谨慎
develop programs in parallel languages even if they are to be run on sequential computers. In
doing this the added information with respect to concurrency and data dependencies would
allow a sequential computer to improve its performance by instruction prefetching, data
caching, and so forth. [9]
To this end research into parallel systems has focused on developing a cost effective hardware
platform such as Beowulf, and cost effective message passing platforms. These messagepassing platforms have been available freely for commercial operating systems, however do
not parallelise a serial problem, and great effort is required by the programmer to do this.
Parallel programs can be expected to have a lifespan of decades, while the platforms on which
they execute have life spans of much less than a decade. Accordingly, such software must be
designed to run on more than one computer over its lifetime.
Whilst there are many parallel languages available for parallel programming, not all are
discussed here. Additionally it is preferable for the programming systems software to utilize
the lower layer of middleware (i.e. a standard message passing library) as described above.
The reason for this is two fold:
1.
Software needs to be portable – An overall goal of parallel and distributed systems
is to ensure that when a piece of software or an application is written that it has
portability, i.e. the ability to execute on many different machines. (As discussed in
Section 3.4).
2.
A programming system must have the ability to abstract and hide most of the
work involved in parallel development and execution. Without this feature, the switch
to parallelisation of programming will be arduous.
Skillicorn [37] argues that parallel programming models and languages should:
1.
容易编程
2.
软件开发方法
Cluster Computing
Be easy to program – the exact structure of the executing program should be
inserted by the translation mechanism (compiler and run-time system) rather than by
the programmer. This implies that a model should conceal:
•
Decomposition of a program into parallel threads.
•
Mapping of threads to processors.
•
Communication among threads.
•
Synchronization among threads.
Have a software development methodology – To bridge the gap between the
information provided by the programmer about the semantic structure of the
program, and the detailed structure required to execute it, requires a firm semantic
foundation on which transformation techniques can be built.
Page 59
Revision Version 2.3
Copyright © 2003
3.
Be architecture-independent - so that programs can be migrated from parallel
computer to parallel computer without having to be redeveloped, or indeed modified
in any non-trivial way. This requirement is essential to permit a widespread software
industry for parallel computers as computer architectures have comparatively short
life spans.
4.
Be easy to understand – A model should be easy to understand and to teach, since
otherwise it is impossible to educate existing software developers to use it.
5.
Be efficiently implementable – A model should be efficiently implementable over a
useful variety of parallel architectures. Note that efficiently implementable should
not be taken to mean that implementations extract every ounce of performance out of
the target architecture.
6.
Provide accurate information about the cost of programs – Any program's design
is driven, more or less explicitly, by performance concerns. Execution time is the
most important of these, but others such as processor utilization or even cost of
development are also important.
独立于结构
容易理解
有效实现
提供编程成本的有效信息
These criteria reflect the belief that developments in parallel systems must be driven by a
parallel software industry based on portability and efficiency.
Skillcorn evaluated programming models in six categories, depending on the level of
abstraction they provide. Those that are very abstract conceal even the presence of parallelism
at the software level. Such models make software easy to build and port, but efficiency is
usually hard to achieve. At the other end of the spectrum, low-level models make all of the
messy issues of parallel programming explicit (how many threads, how to place them, how to
express communication, and how to schedule communication), so that software is hard to
build and not very portable, but is usually efficient.
Most recent models are near the centre of this spectrum, exploring the best trade-offs between
expressiveness and efficiency. However, there are models that are both abstract and are able to
be implemented efficiently, opening the prospect of parallelism as part of the mainstream of
computing, rather than high-performance computing only.
Efficiency
Abstraction
Figure 3-11 – The Parallel Programming Model Efficiency Abstraction Trade-off
Skillcorn’s six categories are as follows:
Page 60
1.
Nothing Explicit, Parallelism Implicit – Models that abstract from parallelism
completely. Such models describe only the purpose of a program and not how it is to
achieve this purpose. Software developers do not need to know even if the program
they build will execute in parallel. Such models are abstract and relatively simple,
since programs need be no more complex than sequential ones.
2.
Parallelism Explicit, Decomposition Implicit – Models in which parallelism is
made explicit, but decomposition of programs into threads is implicit (and hence so is
mapping, communication, and synchronization). In such models, software developers
are aware that parallelism will be used, and must have expressed the potential for it in
programs, but do not know even how much parallelism will actually be applied at
Cluster Computing
Revision Version 2.3
Copyright © 2003
run-time. Such models often require programs to express the maximal parallelism
present in the algorithm, and then reduce that degree of parallelism to the target
architecture, at the same time working out the implications for mapping,
communication, and synchronization.
3.
Decomposition Explicit, Mapping Implicit – Models in which parallelism and
decomposition must both be made explicit, but mapping, communication, and
synchronization are implicit. Such models require decisions about the breaking up of
available work into pieces to be made, but they relieve the software developer of the
implications of such decisions. (i.e. BSP)
4.
Mapping Explicit, Communication Implicit – Models in which parallelism,
decomposition, and mapping are explicit, but communication and synchronization are
implicit. Here the software developer must not only break the work up into pieces, but
must also consider how best to place the pieces on the target processor. Since locality
will often have a marked effect on communication performance, this almost
inevitably requires an awareness of the target processor's interconnection network. It
becomes very hard to make such software portable across different architectures.
5.
Communication Explicit, Synchronization Implicit – Models in which parallelism,
decomposition, mapping, and communication are explicit, but synchronization is
implicit. Here the software developer is making almost all of the implementation
decisions, except that fine-scale timing decisions are avoided by having the system
deal with synchronization.
6.
Everything Explicit – Models in which everything is explicit. Here software
developers must specify all of the detail of the implementation. As noted earlier, a
programming model needs the ability to abstract, as it is extremely difficult to build
software using such models, because both correctness and performance can only be
achieved by attention to vast numbers of details. (i.e. PVM, MPI)
Two high-level applications development packages, BSP and ARCH are briefly discussed
below. Further information on the design and implementation of parallel programs can be
obtained from the online book Designing and Building Parallel Programs. [42]
3.4.2.1.
BSP
Bulk Synchronous Parallelism (BSP) is a parallel programming model that abstracts from lowlevel program structures in favour of supersteps. A superstep consists of a set of independent
local computations, followed by a global communication phase and a barrier synchronization.
Structuring programs in this way enables their costs to be accurately determined from a few
simple architectural parameters, namely the permeability of the communication network to
uniformly-random traffic and the time to synchronize. Although permutation routing and
barrier synchronizations are widely regarded as inherently expensive, this is not the case. As a
result, the structure imposed by BSP comes free in performance terms, while bringing
considerable benefits from an application-building perspective. [25]
3.4.2.2.
ARCH
ARCH is an extension to MPI relying on a small set of programming abstractions that allow
the writing of multi-threaded parallel codes according to the object-oriented programming
style. ARCH has been written on top of MPI with C++. Very little knowledge of MPI is
required to use ARCH as the latter entirely redefines the user interface in an object-oriented
style and with new thread-compatible semantics. C++ was not simply used as a development
language. Instead, it was attempted to transmit the object-oriented method for program
development.
ARCH consists of several sets of C++ classes supplying tools for the writing of multi-threaded
parallel codes.
Cluster Computing
Page 61
Revision Version 2.3
Copyright © 2003
The first set deals with threading and supplies two classes to this purpose: the Thread and
S_Thread classes. The Thread class provides functions for thread construction, destruction,
initialisation, scheduling, suspension and so forth. S_thread is defined by private derivation
from the previous one and allows the writing of multi-threaded programs in a structured style.
Illustration cases may be found in [35] where each multi-threaded program is presented in a
layout similar to an electronic board design. The library contains three additional sets of
classes for thread synchronization and communication. Each set relates to a well-identified
communication model:
Page 62
1.
Point-to-point synchronous.
2.
Point-to-point asynchronous.
3.
One-sided via global write/read function calls.
Cluster Computing
Revision Version 2.3
Copyright © 2003
4. System Installation & Testing
4.1. Building a Beowulf
Whilst this section is not intended to be a complete or comprehensive guide to building a
Beowulf Cluster, it does list the salient points on the configuration process as well as the
explicit configuration used for testing in the laboratory.
Building a Beowulf cluster requires a thorough working knowledge of the operating system
including networking, file systems, system configuration, daemons and services. In relation to
Linux this requires a working knowledge of up to 30 different configuration files and their
individual format requirements to get a system up and running. This detailed configuration
knowledge is learned through reading and implementing (by trial and error) the various HowTos.
With respect to my knowledge and experience with Linux, it dates back to Redhat Linux 5.2
(October 1998) and the documentation that is particularly useful for building a cluster is noted
in [11] and [38.1]
Linux Installation
Initially the Redhat 6.2 and 7.1 (codename: Seawolf, Linux Kernel Version 2.4.7-2)
distributions were initially tested. It was found that both contained many bugs, especially to do
with NFS and EXT2.
As soon as the Redhat 7.2 distribution was available, all nodes in the cluster were upgraded to
RH7.2 as it implements the 2.4.x Kernel and the EXT3 file system that offers greater stability
and performance. The RH7.2 distribution also includes versions of LAM and PVM that can be
installed as part of the initial installations.
For laboratory-testing purposes the Linux Redhat 7.2 distribution will be used implementing
the 2.4.7-10 kernel.
It is of note that the 2.4.x series kernels do not yet implement NFS over TCP instead UDP is
used (which Linux uses by default as opposed to Sun Solaris using TCP).
In total eight nodes were used, the specifications of which is shown in Figure 4-1
Figure 4-1 – Layered Model of Beowulf Cluster Computer Used in Testing
Cluster Computing
Page 63
Revision Version 2.3
Copyright © 2003
In addition to the configuration as detailed in Figure 4-1, the master server was installed with
128Mb of Ram, an additional network card for external cluster access, and a 2Gb hard-drive
for the /home directory which will be cross mounted within the cluster.
Installation
The following process was under taken in installing the system:
1.
Install CD-ROM on the master node (server node).
2.
Install Linux Redhat 7.2 on the master node from the CD-ROM using a CD-ROM
boot image floppy (use the dos utility rawrite from CD-ROM 1).
3.
Configure using Linuxconf and start the required services using serviceconf. (Note: if
serviceconf is not available then symbolic linking is required to the Run Level 3 and 5
startups – however I prefer to use a GUI for system services operations, hence
serviceconf was used).
4.
Create the following accounts to be used throughout the whole cluster:
Name
Password
Purpose
root
cluster
# for administration
beowulf
beowulf
# cluster operation
The root account is local to each machine, however the beowulf account is global
(using the same group and user ID) and hence a change on any node to this account
will be instantaneously reflected across all nodes. Additionally the beowulf account
and home directory are located under /home/beowulf which is exported from the
NFS server, hence only one set of config files exist for this account. The beowulf
account was configured using the csh (C Shell) for compatibility and ease of
configuration (requires conf files for POV-Ray, LAM and other parallel applications).
5.
Start NFS server, exporting /home and /mnt/cdrom to all nodes in the cluster.
6.
Ensure NFS, RSH, NTP, TCP/IP services are working. Use rpcinfo –p to make
sure the NFS server is working (rpc.portmap, rpc.mountd, rpc.nfsd, rpc.statd,
rpc.lockd, rpc.rquotad should be listed).
7.
Copy Linux Redhat 7.2 CDs (1 and 2) from the CD-Rom to /home.
8.
Install Linux Redhat 7.2 on each node over the network using an NFS image and a
network boot image floppy – do not add accounts other than the root account at this
stage.
9.
Configure using Linuxconf and start the required services using serviceconf (all
required files can be configured with these utilities or manually. Using these utilities
is through recommended, however knowledge of the file formats is still required).
10. Add all nodes to the etc/hosts file, such as:
node1
192.168.0.1 # master server, NFS server and /home dir
node2
192.168.0.2
node3
192.168.0.3
node4
192.168.0.4
node5
192.168.0.5
node6
192.168.0.6
node7
192.168.0.7
node8
192.168.0.8 # backup LAM server
11. DNS and NIS were not used to reduce daemon overhead. However could be used in
future clusters to ease of maintenance and management.
Page 64
Cluster Computing
Revision Version 2.3
Copyright © 2003
12. Get Telnet up and running, as well as Linuxconf web access services for remote
management of all nodes. Linuxconf web access allows remote configuration through
a web-browser via port 98. For example:
http://192.168.0.x:98
13. Edit all Beowulf clients /etc/ntp.conf to add the server address
192.168.0.1 and set the Beowulf master clock to the correct time.
14. Cross mount /home and mnt/cdrom from the NFS server on each node of the
system.
15. Remove internal passwords for each node in the cluster using a .rhosts file in the
top level beowulf account directory.
16. Ensure rsh and rlogin access from the server to each node is possible.
17. Stop all unused daemons running on every node in the cluster. This includes lpd and
sendmail for printer and email services respectively.
18. Stop all applications that consume resources such as processing power and memory.
Boot each node into Run-level 3 (Multi-User/Text/Full Network).
19. Install LAM, if not installed in the initial installation. Configure the
/etc/lam/lam-bhost.lam file to contain the node names on the system.
20. Login to each node using a user account, rather than the root account. LAM does
not allow the use of the root account as it could crash and destroy the system (for
testing purposes, it is actually more convenient to log into each node as root and only
the server node as beowulf, the server logs into each node using the beowulf
operating account, even though passwords are not required).
21. Start LAM using the lamboot command on the server [Refer to 8.5.1 for details].
Notes on the Installation:
The following notes are for historical reference purposes:
1.
Linux’s default implementation of NFS is over unreliable data transport UDP rather
than reliable data transport such as TCP. This is in contrast to Sun Solaris. Linux can
be configured to use TCP however the Linux 2.4.x series kernels do not provide
support for this at the time of writing. Should reliability be a design requirement then
the 2.2.x series kernels should be used such as Redhat 6.2. Another benefit of using
earlier distributions is that they have smaller size-footprint. [41]
2.
Whilst the test installation was carried out using a network installation, nodes can be
cloned which reduces the overall work involving in setting up or in adding additional
nodes in a large cluster. Information on cloning nodes can be obtained at:
ftp://ftp.sci.usq.edu.au/pub/jacek/beowulf-utils/disk-less/ [11]
3.
Whilst PVM was not recommended in this document to be used as the lower layer of
middleware, it does provide some noteworthy features such as fault-tolerance and an
easy to use GUI that can be installed with Redhat 7.2.
All nodes in the cluster were located at the University of Technology in the ITS research
laboratory in Building 1, Level 21 room 1/2122E. The Ethernet switch was located in room
1/2122A. As can be seen in Figure 4-2 below, all machines have monitors, keyboards and
mouses, an attribute that is not required in a Beowulf, cluster. Although not required, this does
make configuration, in various test environments easier.
Cluster Computing
Page 65
Revision Version 2.3
Copyright © 2003
Figure 4-2 – Test Cluster Computer in Lab B1/2122E
4.2. Performance Testing
The test suites as outlined in this section were installed and run on the test-cluster.
4.2.1.
Beowulf Performance Suite
The Beowulf Performance Suite (BPS) was developed by Paralogic [38.2] as a graphical front
end to a series of commonly used performance tests for Parallel Computers. It was designed
explicitly for the Beowulf Class of computers running Linux and the suite is packaged as an
easy to install .rpm file. The suite contains seven different but well-known tests, which were
commonly installed separately.
The suite can be run from the command line by invoking bps or xbps with the GUI as shown
in Figure 4-3 below. The suite generates all test results and graphs (using Ggnuplot) in .html
format hence documentation can be maintained on line, which is in-line with the use of webbased configuration and reporting mechanisms to streamline system administration (such as
CISCO Ethernet switch management and node management with linuxconf web access).
Figure 4-3 – BPS Running on node1 of the Test Cluster
Page 66
Cluster Computing
Revision Version 2.3
Copyright © 2003
Details of each of the seven tests are noted below:
硬盘与文件系统
内存带宽
网络性能
1) Bonnie++ – is a benchmark suite that is aimed at performing a number of simple
tests of hard drive and file system performance.
2) Stream – The STREAM Benchmark is the de-facto industry standard benchmark
for the measurement of computer memory bandwidth. The STREAM benchmark
measures "real world" bandwidth sustainable from ordinary user programs - not
the theoretical "peak bandwidth".
3) NetPerf – Netperf is a benchmark that can be used to measure the performance
of many different types of networking. It provides tests for both unidirectional
throughput, and end-to-end latency. The environments currently measurable by
netperf include:
•
TCP and UDP via BSD Sockets
•
DLPI
•
Unix Domain Sockets
•
Fore ATM API
•
HP HiPPI Link Level Access
4) LMBench – is a benchmark suite that tests system bandwidth and latency:
系统带宽与延迟
网络性能
并行处理效率
•
Bandwidth benchmarks (Cached file read, Memory copy, Memory
read/write, Pipe, TCP)
•
Latency benchmarks (Context switching, Networking: connection
establishment, pipe, TCP, UDP, and RPC hot potato, File system
creates & deletes, Process creation, Signal handling, System call
overhead, Memory read latency, Miscellaneous, Processor clock rate
calculation)
5) NetPipe – NetPIPE (A Network Protocol Independent Performance Evaluator) is
a protocol independent performance tool that encapsulates the best of ttcp and
netperf and visually represents the network performance under a variety of
conditions. By taking the end-to-end application view of a network, NetPIPE
clearly shows the overhead associated with different protocol layers. Netpipe
assists with identifying ways of optimising a communication channel to boost the
overall system performance.
6) NAS – The NAS Parallel Benchmarks (NPB) are a set of eight programs
designed to assist in the performance evaluation of parallel supercomputers. The
benchmarks, which are derived from computational fluid dynamics (CFD)
applications, consist of five kernels and three pseudo-applications.
I/O与多任务性能
7) UnixBench – tests the file I/O and kernel multitasking performance resulting in a
system index.
Package Dependencies: Gnuplot (graph plotting tools)
4.2.2.
The Linpack Benchmark
Each of the top 500 supercomputers is measured by the LINPACK benchmark taking the `best'
case performance [29.1]. LINPACK was chosen as it is widely used and performance numbers
are available for almost all relevant systems and is hence a relevant benchmark for clusters.
The LINPACK Benchmark was introduced by Jack Dongarra. The benchmark used in the
LINPACK Benchmark is to solve a dense system of linear equations.
Package Dependencies: blas (basic linear algebra sub-programs)
Cluster Computing
Page 67
Revision Version 2.3
Copyright © 2003
4.3. System Administration
The following system administration tools were used to manage and monitor the test-cluster in
real-time.
4.3.1.
General
Packaged with the RH7.2 distribution are two useful operating system level tools that were
widely used to set-up and manage the workstations:
1.
Linuxconf – Linuxconf is a utility similar to Windows NT control panel. The
majority of settings for the local workstation or server, network configuration, file
systems and user accounts are available to be set in a GUI environment. This is in
stark contrast to the norm of configuring 30 different text files, each with there own
different formats. Another benefit for clustering and networking in general is that
nodes can be managed by Linuxconf over a network through a web browser.
2.
Serviceconf – Serviceconf is a utility again similar to Windows NT service manager.
Serviceconf allows the various services to be started, stoped and set-up for each runlevel in a GUI environment. Within a cluster environment, this is important as there
are many services that need to be run from the xinetd service. Whilst Linuxconf
does have the ability to manage services, it does not extend to the sub-protocols of
xinetd.
A third package BCS (basic cluster scripts) is a package of specially developed parallel tools
[38] and was added for cluster administration purposes. The scripts are implemented in the
Shell BASH, which was used for all root accounts on the cluster.
The scripts were designed so that any group of Linux machines with rcp and rsh can be
acted upon as a group. This assists configuration, i.e. copying a configuration file to each
machine or installing an .rpm across each machine.
4.3.2.
Mosixview
Mosxiview is an extensible management and monitoring tool for Beowulf Clusters, developed
to simplify the management tasks for a large Beowulf Cluster. [1.1]
Real-time monitoring is critical in a clustering environment, as we need a way to ensure that
each node is up and running, has enough resources to run, and is not overloaded.
As shown in Figure 4-4 below, the main window of Mosixview lists all the vital statics of each
machine. Moxixview additionally has a service control window for managing services on
remote nodes. The services manager can start, stop, and allow remote execution of services.
Package Dependencies: QT 2.3.0 (window development libraries)
Figure 4-4 – Main Window of Mosixview Cluster Management Software
Page 68
Cluster Computing
Revision Version 2.3
Copyright © 2003
4.4. Applications Testing
4.4.1.
Persistence of Vision
Persistence of Vision Ray-tracer or POV-Ray is a ray-tracing program that creates threedimensional, photo-realistic images and animations using the ray tracing rendering technique.
It reads in a text file denoted with extension .pov, containing information describing the
objects and lighting in a scene and generates an image of that scene from the viewpoint of a
camera also described in the text file.
Ray-tracing is a rendering technique that calculates an image of a scene by simulating the way
rays of light travel in the real world. However, it does its job backwards by starting with a
simulated camera and traces rays backwards out into the scene. It is done backwards as the
vast majority of rays never hit an observer, hence it would take forever to trace a scene. Ray
tracing is not a fast process by any means, but it produces very high quality images with
realistic reflections, shading, perspective, and other effects as shown in Figure 4-6.
Standard Package Dependencies: libpng (library functions for PNG image manipulation)
Animation rendering is an excellent example of process-level parallelism and as POV-Ray is
freely available for Linux (as well as many other platforms), it is an excellent application for
parallel execution and benchmarking. [12] [3.1.5]
In fact, much data is available on benchmarking computer systems with POV-Ray. Shown in
Figure 4-5 is the result of the standard script skyvase.pov [8.4.1], which was introduced by
Andrew Haveland-Robinson as a benchmark for POVRay on many different computer
systems.
Figure 4-5 – Ray Tracing Bench Mark Test Output
Table 8-3 [refer Appendix E] contains a data extract of the performances of various clusters,
similar in power to the test-cluster, on this benchmark.
The test-cluster will be benchmarked using the skyvase.pov as shown in Figure 4-5,
poolballs.pov as shown in Figure 4-7, and tulips.pov as shown in Figure 4-8. [43]
Cluster Computing
Page 69
Revision Version 2.3
Copyright © 2003
Figure 4-6 – Sample Output of Povray’s abilities
Figure 4-7 – Alternate POV-Ray Test Case 1
Page 70
Cluster Computing
Revision Version 2.3
Copyright © 2003
Figure 4-8 – Alternate POV-Ray test Case II
POV-Ray is an open source project, and hence there are many different versions available,
including a wide range of plugins and patches for various compatibility and extensibility.
Some of the plugins available include:
•
Graphical User Interface.
•
Plugin for compatibility with RenderMan.
•
Plugin for compatibility with AutoCAD.
•
Patch for PVM to operate POV-Ray in parallel with PVM.
•
Patch for MPI to operate POV-Ray in parallel with MPI/MPICH.
One of the patches that was not available was explicit support for the LAM implementation of
MPI. Since this was the desired distribution due to performance and compatibility, it was
decided to use the MPICH version of the patch [MPICH/POV-Ray patch by Leon Verrall
Cluster Computing
Page 71
Revision Version 2.3
Copyright © 2003
www.verrall.demon.co.uk/mpipov/] and modify for LAM compatibility. However, it was
found that no modifications were required and compilation of the new version of MPI-POVRay using the existing Linux source code to suit LAM was successfully completed.
Theory of MPI-Povray
Using the new version of MPI-POV-Ray, the system operates with one master task and many
slave tasks. The master has the responsibility of dividing the image into small blocks that are
assigned to slaves. When a slave has finished rendering a block, it is sent back to the master.
The master combines them to form the final image. The master task does not render anything
itself, and hence does not use much CPU power. For better utilization of the master, a slave
task (i.e. rendering) can be run on the master.
MPI POV-Ray adds the following additional commands line options to standard POV-Ray:
+NHxxx Height of MPI Chunks
+NWxxx Width of MPI chunks
The options control the dimensions of the chunks that get assigned to slave tasks. In general,
this allows customisation to suit the particular architecture and system performance that MPIPO-Ray is operating on. In general, it is a trade-off between message passing overhead and
division of labour.
If the chunk size is set too high, it is likely that one PE (processing element) will get a difficult
section of the image to render, and this will be come rate-limiting for the while job. If the
chuck size is too small, much processing time is spent passing messages (or blocking on
Sen/Receive messages), hence under utilizing CPU power.
As a guide if the image is equally difficult to render in all parts of the image and less than 16
processors are available, then a larger chunk size is recommended (try to divide the image
equally between the available PE’s).
If the image varies in difficulty or there are a large number of processors available, then the
default chunk size of 32x32 (in a checker configuration) is recommended. [43] [1]
With animations, although not fully tested in this document, each PE is assigned its own
frame.
Page 72
Cluster Computing
Revision Version 2.3
Copyright © 2003
5. Results
5.1. Summary of Numerical Data
This section details the results of applications testing and includes a results analysis. The raw
results obtained from the test cluster, including scripts of command line input/output can be
found in Appendix E.
From the large amount of data and test results obtained in the laboratory, the most interesting
are detailed in this section (applications level only) to show the various speedup indices of a
cluster computer. Shown in Table 5-1 below, are the results of the rendering test cases as
detailed in section 4.4.1.
skyvase.pov
No. of Render Time
(seconds)
CPUs
1
2
3
4
5
6
7
8
129
72
48
36
29
25
22
20
Speedup
1.0
1.8
2.7
3.6
4.4
5.2
5.9
6.5
poolballs.pov
Render Time Speedup over
(seconds)
X-POV-Ray1
1
2
2541 / 1786
906
605
457
364
309
262
229
1.0
2.8
4.2
5.6
7.0
8.2
9.7
11.1
tulips.pov
Speedup on
MPI-X-POV2
Ray
1.0
2.0
3.0
3.9
4.9
5.8
6.8
7.8
Render Time
(seconds)
Speedup
279264
1.0
Failed
Table 5-1 – Rendering Test Results
5.2. Results Analysis
With respect to Table 5-1 the following analysis of the numerical results is presented:
Skyvase
The skvase.pov test case is the benchmark for a cluster computer using parallel rendering.
The results show that a near ideal speedup was achieved.
From the POVBENCH benchmark data [Table 8-3] it can be seen that for a single node the test
cluster achieved a POVmark of 114.73 which is considered above average for the hardware
used (based on the POVmark data). This can be attributed to the use of structured design and
implementation of Linux on each node as well as efficiencies included in the implemented
Linux kernel version 2.4.7-10.
When rendered on all eight nodes the test cluster achieved a POVmark of 740.00. When
comparing the cluster test results with the POVBENCH cluster data, the test cluster takes the
86th spot with a parallel rendering time of 20sec, out of the 236 parallel results submitted. This
is just below the 85th spot with a 16 node cluster of Pentium Pro 200MHz machines with a
rendering time of 19sec.
These high-performing results can be attributed to the high-performing nodes, a 100BaseTX
switched Ethernet fabric and the use of the high performance MPI-library LAM for POV-Ray
that is not commonly used.
Cluster Computing
Page 73
Revision Version 2.3
Copyright © 2003
Poolballs
Of all of the test-cases the Poolballs test results contained the most interesting trend.
The test case was run under two scenarios, as follows:
1.
X-POV-Ray – Poolballs was first run on one node using the standard serial version
of POV-Ray. This resulted in a best execution time of 2541 seconds. All further
results were obtained using the parallel version as required for multiple CPU’s. It can
be seen from the results that a super-linear speedup was achieved over the serial
version.
2.
MPI-X-POV-Ray – Whilst the first scenario was run many times, each time
showing a super-linear speedup was achieved, it was decided to test the case when
the Parallel version is run on one node using a master and slave process. This
required the operating system to multitask between the two processes and it was
expected that a similar result to the serial version would be achieved. However as
shown in Table 5-1, the parallel version out-performed the serial version on one
node. This changes the speedup results when compared to the parallel version,
showing a close to ideal speedup.
Tulips
The tulips test case has a high degree of difficulty and was used to test the stability of the
cluster. Additionally it was intended to provide a longer test case which could show the trend
as many messages are sent and received over the network, increasing the load on the server,
the communications medium, and each node.
Unfortunately, the test case took over 77 hours to render on one node and as such results were
not obtained for each combination of processors. The test case was run on all eight nodes,
however it was apparent that after five days that it had crashed the cluster, and due to time
constraints further testing to get it running to completion was not possible. Hence, this test case
has not provided any information on speedup index.
From the work that was done using the rendering test cases, in particular Tulips, it is theorized,
that as the resources are heavily utilized the speedup index decreases to the lower bound
Log2n.
Shown in Figure 5-1 below is a plot of the salient test results: skyvase, poolballs super linear
speedup and the applicable theoretical expectations as derived in section 3.1.3.
Page 74
Cluster Computing
Revision Version 2.3
Copyright © 2003
12
10
Speedup
8
6
Ideal Speedup
4
Log2n
Amdahls f = 0.05
Skyvase Speedup
2
PoolBalls Speedup
0
0
2
4
6
8
Number of Processors
Figure 5-1 – Test Results & Theoretical Performance Comparison
It was noted from the raw test results produced by MPI-POV-Ray, that each node was not
contributing an equal percentage of work. Whilst MPI-POV-Ray divides an image up equally,
it does not evaluate the difficulty of each chunk. Each chunk is different and takes a different
time to render. As compensation to this, MPI-POV-Ray distributes small chunks (also chunk
size can be modified by the user to suit the particular architecture) such that the disparity is
minimised. However, this disparity does lead to a decrease in the overall speedup index.
Whilst this is minimal for the rendering algorithm, as it can be considered symmetric, this is
not the case for all algorithms, and applications.
[Refer to Appendix E – 8.5.2: Within each multi-CPU test case are PE (PE) Distribution
Statistics. These statics detail the percentage of work that each node completed].
The following conclusions can be draw from these results:
Cluster Computing
•
The performance of a Beowulf Cluster can be modelled by a pessimistic Lower bound
Log2n for n ≤ 8 and the upper bound, which is the ideal case n.
•
Amdahl’s law with f=0.05 models more closely the lower bound than Log2n
however it is noted that both are highly pessimistic models when trended out to larger
numbers of processors such as 64, 128 and 512.
•
In some instances super-linear speedups can be achieved, however only with treesearch type applications or inherently parallel algorithms.
•
Through the process of structured design, a high-performing, cost-effective Cluster
Computer can be built with known performance bounds.
•
Load Balancing for nodes is generally required. Should this feature not be available it
is possible to manually balance work using the weighted harmonic mean principle [as
detailed in 3.1.3]. The outcome of this principle is that each node should have the
same physical hardware specification and performance. This enables algorithms that
are naturally well balanced (i.e. symmetric such as parallel rendering) to not require a
load-balancing facility to work effectively and achieve a high speedup index. This
methodology can cater effectively for the general-purpose case with a symmetric
algorithm. Whilst rates will invariably become out of balance, even in an application
such as MPI-POV-Ray [43], the imbalance will be minimised.
Page 75
Revision Version 2.3
Copyright © 2003
6. Conclusion
The Beowulf cluster computer is classified as a MIMD, loosely coupled class computer using
the message-passing model for node intercommunication. It is built from commodity class
components enabling it to be cost effective, in comparison to traditional super-computers.
The motivation to build cluster computers can be seen from the industries that use these
systems. They generally require large amounts of cost-effective computing power, to derive
results in a timely fashion.
The Beowulf class of computers were started in 1994 by NASA and have grown immensely
since, which can be seen by the proliferation in the world top500 supercomputers as well as the
available software for the architecture.
Beowulf clusters can be effectively designed to suit different applications depending on the
performance requirements and the commodity hardware available. Cost-savings are introduced
with the use of commodity components available in the open market as well as the many
system choices that can be made (such as Hardware and Operating Systems) rather than
specific vendor solutions.
Beowulf-class computers can be described as:
•
Cost-effective.
•
Easy to assemble.
•
Reliable.
•
Deliver on wide range of carefully chosen applications.
However, Beowulf-class computers are not:
•
Appropriate for all supercomputer applications.
•
Infinitely scalable.
•
Not for all applications and every situation.
•
Not user-intuitive to set-up and use.
This document has shown that the best approach to building high performance clusters is using
a layered approach. Each layer can be looked at individually selecting the highest performing
and most cost-effective solution. The results of this analysis are:
Page 76
•
The Linux operating system has the features and abilities of commercially available
operating systems, in some cases out-performing them with speed and the ability to
reduce the footprint to a minimal size. Linux one of very few operating systems to
offer these features with the added benefit of an open licensing arrangement, reducing
the overall cost of each node for a Beowulf cluster.
•
MPI offers many features for the parallel programmer and functionality as the lowest
layer of middleware. The most notable is the portability between systems and
performance. In the coming years the MPI standard will completely overrun the use
of PVM, when it includes the fault-tolerance, load-balancing and other mature
features it currently lacks. As with Linux, open source and free distributions of MPI
are available for use with Linux clusters. One of the highest performing, with the
highest degree of compatibility with the current version of the MPI standard is the
Local Area Multicomputer (LAM) distribution.
•
Broad testing of the Beowulf cluster is required to optimise the platform and provide
verification of the design objectives. Many test suites are available for clusters,
Cluster Computing
Revision Version 2.3
Copyright © 2003
however applications level testing should be completed such as Parallel rendering
with MPI-POV-Ray.
As part of the work outlined in this document, theoretical performance figures were taken in to
the laboratory for verification. This process included the design, set-up and commissioning of
a high-performance Beowulf cluster using the Linux Operating System, the LAM MPI
message passing library and a pseudo-novel implementation of the ray-tracing rendering and
animation program POV-Ray that was built for LAM/MPI.
During the testing process it was found that in some instances running an application in
parallel can result in high performance, and that parallel applications can gain super-linear
speedups over their serial counterparts. These efficiencies can only be tapped with parallel
systems and for certain applications.
The results show great variability in the performance of an application, based on the input data.
Whilst there is variance in performance, it was shown that it is possible to determine the
bounds of performance prior to implementation, thus leading to a meaningful design process.
Load balancing, the technique of ensuring that each node is processing it fair share of work
and is not under or over utilised, is not necessarily required to produce high-performance
speedups on symmetric algorithms. It is possible through system design and application
implementation to minimise the requirement for load balancing. However, it is a recommended
facility, integral to the middleware layer of software and implemented in packages such as
PVM and in the future MPI. In some instances, it is part of the Operating System (as with the
Cluster Versions of Windows 2000 and Linux Redhat).
The coming parallel software generation will see a myriad of languages utilized, all offering
varying degrees of abstraction and efficiency. In the long term it will be most likely that a few
specialised languages will become dominant in the main stream parallel applications
development market. Efficiency will be compromised for ease of development. In contrast the
high performance scientific community will continue to use and develop specialised scientific,
portable and highly efficient languages for their own use. These languages will have abilities
to tweak applications for different platforms to maximise the performance without code-level
modifications.
The Future of Clusters
Clusters have an ever growing base in the following areas:
•
Scientific and Engineering research (Universities, Government and Industry).
•
Education.
•
Computer Science.
•
Corporations.
•
Media Development Industries.
As the cost of hardware and storage continues to decline, the use of clusters will increase
dramatically. Already most database vendors support SMP machines and several, including
Oracle, are starting to release versions of their software for Linux clusters. As these mainstay
packages become available, they will drive the availability of a whole new class of
applications that run on these machines, ranging from serious business applications to
entertainment applications such as online gaming systems.
Another effect of the ever-increasing capability of computer hardware is its ever-decreasing
size. Traditional supercomputers are very large and until recently, clusters were even larger
since they are, by definition, collections of rack-mounted workstations. As system sizes
decrease, the physical size of a cluster will decrease as well, while the overall computational
capabilities will increase. With the current availability of two nodes in one RU rack space, a
64-node cluster can fit in a single 19” rack, and with single-board computers, the same cluster
is able to fit under a desk.
Cluster Computing
Page 77
Revision Version 2.3
Copyright © 2003
7. References
Other documents referenced within this document:
1. Beowulf How-to
Beowulf UnderGround Website
Version 1.1.1, November 1998
Jacek Radjewski and Douglas Eadline
Available on the Internet at:
http://www.beowulf-underground.org/
1) Mosixview – http://www.beowulf-underground.org/software.html
2. PVM: Parallel Virtual Machine
A Users’ Guide and Tutorial for Networked Parallel Computing
Al Geist, Adam Beguelin, Jack Dongarra, Weicheng Jiang, Robert Manchek,
Vaidy Sunderam
The MIT Press
Cambridge, Massachusetts, London, England
1994 Massachusetts Institute of Technology
Available on the Internet at:
http://www.netlib.org/pvm3/book/pvm-book.html
3. MPI: The Complete Reference
MPI: The Complete Reference
Marc Snir, Steve Otto, Steven Huss-Lederman, David Walker, Jack
Dongarra
The MIT Press
Cambridge, Massachusetts
London, England
4. Beowulf at NASA/GSFC
Beowulf at NASA/GSFC
Earth and Space Sciences Project
NASA Goddard Space Flight Center
http://beowulf.gsfc.nasa.gov/
Website now maintained by Scyld Computing Corporation:
1) http://www.beowulf.org/software/bonding.html – Beowulf Channel
Bonding
2) http://www.beowulf.org/intro.html – Beowulf Introduction
5. Modern Operating Systems
Andrew S. Tanenbaum
Vrije Universiteit
International Edition 1992
Prentice-Hall International, Inc
Page 78
Cluster Computing
Revision Version 2.3
Copyright © 2003
6. Computer Architecture and Parallel Processing
Kai Hwang and Fayĕ A. Briggs
International Edition 1985
McGraw Hill
7. An Introduction to the Intel Family of Microprocessors
A Hands-On Approach Utilzing the 8088 Microprocessor
James L. Antonakos
Second Edition 1996
Prentice Hall
P41 – Pipelining in the Pentium and 486 chip
8. Programming Distributed Systems
Henri Bal
First Edition 1990
Prentice Hall
9. Introduction to Parallel Processing
Algorithms and Architectures
Behrooz Parhami
First Edition 1999
Plenum Press
P15 – 1.4 Types of Parallelism: Taxonomy
P78 – Table 4.2 Topological Parameters of Selected Interconnection Newtorks
10. Pocket Glossary of Computer Terms
Including Technical Reference Material
Black Box Corporation
Third Edition 1999
11. Beowulf Installation and Administration How-to
Beowulf UnderGround Website
Version 0.1.2, June 1999
Jacek Radjewski and Douglas Eadline
Available on the Internet at:
http://www.beowulf-underground.org/
12. How to Build a Beowulf
A Guide to the Implementation and Application of PC Clusters
Thomas L. Sterling, John Salmon, Donald J. Becker, Daniel F. Savarese
First Edition 1999
The MIT Press
Cambridge, Massachusetts
London, England
P11 – Figure 2.1 Classification Scheme for Parallel Computing
Cluster Computing
Page 79
Revision Version 2.3
Copyright © 2003
13. RFC 1918
Address Allocation for Private Internets
Y. Rekhter, B. Moskowitz, D. Karrenberg, G. J. de Groot, E. Lear
Network Working Group
February 1996
Obsoletes RFC: 1627, 1597
BCP: 5 Category: Best Current Practice
Available on the Internet at:
http://www.cis.ohio-state.edu/cgi-bin/rfc/rfc1918.html
14. Parallel Processing
From Applications to Systems
Dan I. Moldovan
First Edition 1993
Morgan Kaufmann
P26 – 1.4 Performance of Parallel Computations
15. Orca
Report on the Programming Language Orca
Henri E. Bal
Dept. of Mathematics and Computer Science
Vrije Universiteit
Amsterdam, The Netherlands
May 1994
16. A Programmers Guide to ZPL
Lawrence Snyder
Department of Computer Science and Engineering
University of Washington
Seattle WA 98195
Version 6.3, January 6, 1999
Available on the Internet at:
http://www.cs.washington.edu/research/zpl/
17. Bulk Synchronous Parallel
BSP Clusters: High performance, Reliable & Very Low Cost
Stephen Donaldson and Jonathan M.D. Hill
Programming Research Group, University of Oxford, U.K.
Questions & Answers About BSP
Stephen Donaldson and Jonathan M.D. Hill
Programming Research Group, University of Oxford, U.K.
18. Logic and Computer Design Fundamentals
M. Morris and Charles R. Kime
International Edition 1997
Prentice Hall
P572 – 12-2 Locality of Reference
Page 80
Cluster Computing
Revision Version 2.3
Copyright © 2003
19. LAM / MPI Parallel Computing
Website http://www.lam-mpi.org/
1) http://www.lam-mpi.org/mpi/mpi_top10.php – MPI vs. PVM
2) http://www.lam-mpi.org/mpi/implementations/ – MPI implementations
3) http://www.lam-mpi.org/linux/ – LAM/Linux issues
4) http://www.lam-mpi.org/performance.php – LAM vs. MPICH
performance comparisons
20. MPI Forum
Message Passing Interface Standards Website
MPI 2 Standard available online
Website http://www.mpi-forum.org/
21. PVM and MPI
A Comparison of Features
G. A. Geist, J. A. Kohl, P. M. Papadopoulos
May 30, 1996
Note this paper compares MPI 1.2 with PVM 3.4
http://www.epm.ornl.gov/pvm/PVMvsMPI.ps
22. Redhat
Commercial and Free Version of Linux
Software Website
http://www.redhat.com/
23. Intel
Corporate Website
http://www.intel.com/
2) http://www.intel.com/eBusiness/products/enterprise/8way/8wayQRG.htm –
SMP Block diagram
24. Los Alamos National Laboratory
Centre for Non-Linear Studies
Avalon Beowulf Cluster
http://cnls.lanl.gov/Internal/Computing/Avalon
25. University of Oxford Parallel Applications Centre
The Bulk Synchronous Parallel Model
University Website:
http://oldwww.comlab.ox.ac.uk/oucl/oxpara/bsp/bspmodel.htm
26. University of California NOW Project
Conducted by the Computer Science Division
University of California, Berkeley
University Website: http://now.cs.berkeley.edu/
27. Scyld Computing Corporation
Corporate Website
http://www.scyld.com/
Cluster Computing
Page 81
Revision Version 2.3
Copyright © 2003
28. A System Software Architecture for High-End Computing
David S. Greenberg et al
Sandia National Laboratories
http://www.supercomp.org/sc97/proceedings/TECH/GREENBER/INDEX.HTM
29. Top 500 Supercomputer Sites
Benchmark performance and ranking the of the top 500 Supercomputers
University of Mannheim
University of Tennessee
All information was extracted from the November 2001 Top 500 List
http://www.top500.org/
1) http://www.top500.org/lists/linpack.html - Linpack Config
2) http://www.top500clusters.org/ - Top 500 Clusters
30. Linux Parallel Processing How-to
Purdue University
Version 980105, 5 January 1998
Hank Dietz
Available on the Internet at:
http://yara.ecn.purdue.edu/pplinux/
31. High Performance Cluster Computing
Volume 1 – Architectures & Systems
Volume 2 – Programming & Applications
Edited By Rajkumar Buyya
Monash University Melbourne, Australia
First Edition 1999
Prentice Hall PTR
P10 – Figure 1.2 Cluster Computer Architecture
32. Operating Systems
Internals and Design Principles
William Stallings
Fourth International Edition 2001
Prentice-Hall International, Inc
33. Electronic Design
Technology-Applications-Products-Solutions
http://www.elecdesign.com/
Penton
August 17, 1998 Issue
Page 30 - Mail-Order Supercomputer available for a mere $150,000
34. Supporting Microsoft Cluster Server
Microsoft Official Curriculum
CD-ROM Version
December 1, 1997
Page 82
Cluster Computing
Revision Version 2.3
Copyright © 2003
35. Multi-Threaded Object-Oriented MPI-Based Message Passing
The ARCH Library
Jean-Marc Adamo
First Edition 1998
Kluwer Academic Publishers
http://www.cpe.fr/~arch/ARCH-index.htm
Chapter 8 – Parallel A* Algorithm
36. Parallel Algorithms/Architecture Synthesis
The Second Aizu International Symposium
March 17-21 Japan
IEEE Computer Society Press
Page 340 – A Parallel & Fault-Tolerant LAN with Dual Communications
Subnetworks
37. Models and Languages for Parallel Computation
David B. Skillicorn & Domenico Talia
October 1996
38. Paralogic Inc.
Corporate Website
http://www.plogic.com/index.html
http://www.xtreme-machines.com/x-cluster-qs.html - Cluster Quick Start
ftp://ftp.plogic.com - Ethernet Channel Bonding Kernel Patch
1) Paper: Performance Considerations for I/O Dominant Applications on
Parallel Computers
2) Beowulf Performance Suite – http://www.plogic.com/bps
39. GNU General Public License
Version 2, June 1991
Copies of the license can be down found at
http://www.fsf.org/copyleft/gpl.html
40. The Parallel Virtual File System
Research Website
Parallel Architecture Research Laboratory
Clemson University, Clemson, South Carolina
http://parlweb.parl.clemson.edu/pvfs/index.html
41. Linux NFS How-to
28 December 2001
Tavis Barr, Nicolai Langfeldt, Seth Vidal
Available on the Internet at:
http://nfs.sourceforge.net
42. Designing and Building Parallel Programs
28 December 2001
Ian Foster
Online book available on the Internet at:
http://www-unix.mcs.anl.gov/dbpp/
Cluster Computing
Page 83
Revision Version 2.3
Copyright © 2003
43. Persistence of Vision Ray-tracing
1) Main website – http://www.povray.org/
2) Benchmarking – http://www.haveland.com/index.htm?povbench/index.htm
3) Examples inc .pov files – http://www.xlcus.com/povray/
All documentation is available (depending on issue status) upon request via e-mail from
Richard Morrison ([email protected]) quoting the document name and number.
Page 84
Cluster Computing
Revision Version 2.3
Copyright © 2003
8. Appendix
This section of the document contains the appendices that support the work presented in the
main document.
Cluster Computing
Page 85
Revision Version 2.3
Copyright © 2003
8.1. Appendix
A
Technologies
–
Node
Interconnection
This appendix details commonly used networking technologies [30] used to interconnect
cluster nodes. It specifically contains the information required to evaluate, design and
implement a Beowulf cluster network for a specific requirement. This section is divided into
two categories as follows:
1
•
Class 1 – Commodity components (for use in Beowulf Class 1 Clusters)
•
Class 2 – Vendor specific components (for use in Beowulf Class 1 Clusters)
Denotes at the time of writing the information has not been confirmed.
8.1.1.
Class 1 Network Hardware
CAPERS
CAPERS (Cable Adapter for Parallel Execution and Rapid Synchronization) is a spin-off of
the PAPERS project, http://garage.ecn.purdue.edu/~papers/, at the Purdue University School of
Electrical and Computer Engineering. In essence, it defines a software protocol for using an
ordinary "LapLink” parallel SPP-to-SPP cable to implement the PAPERS library for two
Linux PCs. The network do not scale, but the price cannot be beaten. As with TTL_PAPERS,
to improve system security, there is a minor kernel patch recommended (but not required) and
is available at: http://garage.ecn.purdue.edu/~papers/giveioperm.html
Design Considerations:
Linux support: AFAPI library
Maximum bandwidth: 1.2 Mb/s
Minimum latency: 3 microseconds
Available as: commodity hardware
Interface port/bus used: SPP
Network structure: cable between 2 machines
Cost per machine connected: $2
Ethernet
For some years now, 10 Mbits/s Ethernet has been the standard network technology. Good
Ethernet interface cards can be purchased for well under $100, and a fair number of PCs now
have an Ethernet controller built-into the motherboard. For lightly-used networks, Ethernet
connections can be organized as a multi-tap bus without a hub; such configurations can serve
up to 200 machines with minimal cost, but are not appropriate for parallel processing. Adding
an unswitched hub does not really help performance. However, switched hubs that can provide
full bandwidth to simultaneous connections cost only about $150 per port. Linux supports an
extensive range of Ethernet interfaces, but it is important to consider that variations in the
interface hardware can yield significant performance differences. See the Linux Hardware
Compatibility HOWTO for comments on which are supported and how well they work; also
see http://cesdis1.gsfc.nasa.gov/linux/drivers/.
An interesting way to improve performance is offered by the 16-machine Linux cluster work
done in the Beowulf project, http://cesdis.gsfc.nasa.gov/linux/beowulf/beowulf.html, at NASA
CESDIS. There, Donald Becker, who is the author of many Ethernet card drivers, has
Page 86
Cluster Computing
Revision Version 2.3
Copyright © 2003
developed support for load sharing across multiple Ethernet networks that shadow each other
(i.e., share the same network addresses). This load sharing is built-into the standard Linux
distribution, and is done invisibly below the socket operation level. Because hub cost is
significant, having each machine connected to two or more hubless or unswitched hub
Ethernet networks can be a very cost-effective way to improve performance. In fact, in
situations where one machine is the network performance bottleneck, load sharing using
shadow networks works much better than using a single switched hub network.
Design Considerations:
Linux support: kernel drivers
Maximum bandwidth: 10 Mb/s
Minimum latency: 100 microseconds
Available as: Commodity hardware
Interface port/bus used: PCI
Network structure: switched or unswitched hubs, or hubless bus
Cost per machine connected: $150 (hubless, $100)
Ethernet (Fast Ethernet)
Fast Ethernet refers to a hub-based 100 Mbits/s Ethernet that is somewhat compatible with
older "10 BaseT" 10 Mbits/s devices and cables. As might be expected, anything called
Ethernet is generally priced for a volume market, and these interfaces are generally a small
fraction of the price of 155 Mbits/s ATM cards. A drawback is that having a collection of
machines dividing the bandwidth of a single 100 Mbits/s "bus" (using an unswitched hub)
yields performance that is not as good on average as using 10 Mbits/s Ethernet with a switched
hub that can give each machine's connection a full 10 Mbits/s.
Switched hubs that can provide 100 Mbits/s for each machine simultaneously are expensive,
but prices are dropping rapidly per quater, and these switches do yield much higher total
network bandwidth than unswitched hubs. The thing that makes ATM switches so expensive is
that they must switch for each (relatively short) ATM cell; some Fast Ethernet switches take
advantage of the expected lower switching frequency by using techniques that may have low
latency through the switch, but take multiple milliseconds to change the switch path. Should
the design routing pattern be expected to change frequently, then these switches should be
avoided. See http://cesdis1.gsfc.nasa.gov/linux/drivers/index.html for information about the
various cards and drivers.
Channel Bonding is also available as described in Section 3.2.3.1.
Design Considerations:
Linux support: kernel drivers
Maximum bandwidth: 100 Mb/s
Minimum latency: 80 microseconds
Available as: commodity hardware
Interface port/bus used: PCI
Network structure: switched or unswitched hubs
Cost per machine connected: $4001
Cluster Computing
Page 87
Revision Version 2.3
Copyright © 2003
Ethernet (Gigabit Ethernet)
Gigabit Ethernet, http://www.gigabit-ethernet.org/, does not have a good technological reason
to be called Ethernet. However the name does accurately reflect the fact that this is intended to
be a cost effective, mass-market, computer network technology with native support for IP.
However, current pricing reflects the fact that Gb/s hardware is still complex hardware to
manufacture.
Unlike other Ethernet technologies, Gigabit Ethernet provides for a level of flow control that
should make it a more reliable network. Full-Duplex Repeaters (FDRs) or, simply multiplex
lines, using buffering and localized flow control to improve performance. Most switched hubs
are being built as new interface modules for existing gigabit-capable switch fabrics.
Switch/FDR products have been shipped or announced by the majority of networking vendors
such as:
•
http://www.acacianet.com/
•
http://www.gigalabs.com/
•
http://www.baynetworks.com/
•
http://www.packetengines.com/
•
http://www.cabletron.com/
•
http://www.plaintree.com/
•
http://www.networks.digital.com/
•
http://www.prominet.com/
•
http://www.extremenetworks.com/
•
http://www.sun.com/
•
http://www.foundrynet.com/
•
http://www.xlnt.com/
NIC’s and Switches are now available using a Category 5e-cabling infrastructure, rather than
requiring the use of optical fibre. This is significantly driving the cost down as copper cabling
is less expensive and existing copper-cable plants can be reused.
There is a Linux driver, http://cesdis.gsfc.nasa.gov/linux/drivers/yellowfin.html, for the Packet
Engines "Yellowfin" G-NIC, http://www.packetengines.com/. Early tests under Linux
achieved about 2.5x higher bandwidth than could be achieved with the best 100 Mb/s Fast
Ethernet; with gigabit networks, careful tuning of PCI bus use is a critical factor. There is little
doubt that driver improvements, and Linux drivers for other NICs, will follow.
Design Considerations:
Linux support: kernel drivers
Maximum bandwidth: 1,000 Mb/s
Minimum latency: 300 microseconds1
Available as: multiple-vendor hardware
Interface port/bus used: PCI
Network structure: switched hubs or FDRs
Cost per machine connected: $2,5001
Page 88
Cluster Computing
Revision Version 2.3
Copyright © 2003
PLIP
For the minimal cost of a "LapLink" cable, PLIP (Parallel Line Interface Protocol) allows two
Linux machines to communicate through standard parallel ports using standard socket-based
software. In terms of bandwidth, latency, and scalability, this is not a very serious network
technology; however, the near-zero cost and the software compatibility are useful. The driver
is part of standard Linux kernel distributions.
Design Considerations:
Linux support: kernel driver
Maximum bandwidth: 1.2 Mb/s
Minimum latency: 1,000 microseconds?
Available as: commodity hardware
Interface port/bus used: SPP
Network structure: cable between 2 machines
Cost per machine connected: $2
SLIP
Although SLIP (Serial Line Interface Protocol) is firmly planted at the low end of the
performance spectrum, SLIP (or CSLIP or PPP) allows two machines to perform socket
communication via ordinary RS232 serial ports. The RS232 ports can be connected using a
null-modem RS232 serial cable, or they can even be connected via dial-up through a modem.
In any case, latency is high and bandwidth is low, so SLIP should be used only when no other
alternatives are available. It is worth noting, however, that most PCs have two RS232 ports, so
it would be possible to network a group of machines simply by connecting the machines as a
linear array or as a ring. There is also load-sharing software called EQL.
Design Considerations:
Linux support: kernel drivers
Maximum bandwidth: 0.1 Mb/s
Minimum latency: 1,000 microseconds1
Available as: commodity hardware
Interface port/bus used: RS232C
Network structure: cable between 2 machines
Cost per machine connected: $2
Cluster Computing
Page 89
Revision Version 2.3
Copyright © 2003
USB (Universal Serial Bus)
USB (Universal Serial Bus, http://www.usb.org/) is a hot-pluggable conventional-Ethernetspeed, bus for up to 127 peripherals ranging from keyboards to video conferencing cameras.
Computers are interconnected using a USB hub (similar to Ethernet) or interconnected to the
adjacent node(s) (PC’s generally come with two USB ports, these can be split to interconnect
more than two nodes if required).
USB is considered the low-performance, zero-cost, version of FireWire.
USB ports are a defacto standard on PC motherboards similar to as RS232 and SPP.
Development of a Linux driver is discussed at http://peloncho.fis.ucm.es/~inaky/USB.html
Design Considerations:
Linux support: kernel driver
Maximum bandwidth: 12 Mb/s
Minimum latency: Data not available
Available as: commodity hardware
Interface port/bus used: USB
Network structure: bus
Cost per machine connected: $5 – depending on network topology
Page 90
Cluster Computing
Revision Version 2.3
8.1.2.
Copyright © 2003
Class 2 Network Hardware
Myrinet
Myrinet http://www.myri.com/ is a LAN designed to also serve as a ‘System Area Network’
(SAN), i.e., the network within a cabinet full of machines connected as a parallel system. The
LAN and SAN versions use different physical media and have different characteristics;
generally it is recommended to use the SAN version within a cluster.
Myrinet is conventional in structure and has a reputation for being a particularly good network
implementation. The drivers for Linux are said to perform very well, although large
performance variations have been reported with different PCI bus implementations for the host
computers.
Currently, Myrinet is a preferred network technology by cluster-groups (as it has one of the
smallest latencies) that do not have significant budgetary constraints.
Design Considerations:
Linux support: Library
Maximum bandwidth: 1,280 Mb/s
Minimum latency: 9 microseconds
Available as: single-vendor hardware
Interface port/bus used: PCI
Network structure: Switched hubs
Cost per machine connected: $1,800
Parastation
The ParaStation project http://wwwipd.ira.uka.de/parastation at University of Karlsruhe
Department of Informatics is building a PVM-compatible custom low-latency network. They
first constructed a two-processor ParaPC prototype using a custom EISA card interface and
PCs running BSD UNIX, and then built larger clusters using DEC Alphas. Since January 1997,
ParaStation has been available for Linux. The PCI cards are being made in cooperation with a
company called Hitex (see http://www.hitex.com/parastation/). Parastation hardware
implements both fast, reliable, message transmission and simple barrier synchronization.
Design Considerations:
Linux support: HAL or socket library
Maximum bandwidth: 125 Mb/s
Minimum latency: 2 microseconds
Available as: single-vendor hardware
Interface port/bus used: PCI
Network structure: Hubless mesh
Cost per machine connected: > $1,000
Cluster Computing
Page 91
Revision Version 2.3
Copyright © 2003
ArcNet
ARCNET is a local area network that is primarily intended for use in embedded real-time
control systems. Like Ethernet, the network is physically organized either as taps on a bus or
one or more hubs, however, unlike Ethernet, it uses a token-based protocol logically
structuring the network as a ring. Packet headers are small (3 or 4 bytes) and messages can
carry as little as a single byte of data. Thus, ARCNET yields more consistent performance than
Ethernet as it has bounded delays. Unfortunately, it is slower than Ethernet and less popular,
making it more expensive. More information is available from the ARCNET Trade
Association at http://www.arcnet.com/
Design Considerations:
Linux support: kernel drivers
Maximum bandwidth: 2.5 Mb/s
Minimum latency: 1,000 microseconds1
Available as: multiple-vendor hardware
Interface port/bus used: ISA
Network structure: unswitched hub or bus (logical ring)
Cost per machine connected: $200
ATM
ATM (Asynchronous Transfer Mode) has been heralded over the past few years to take over
Ethernet, however this is unlikely to happen. ATM is cheaper than HiPPI and faster than Fast
Ethernet, and it can be used over the very long distances and as such is used in Telco
applications. The ATM network protocol is also designed to provide a lower-overhead
software interface and to more efficiently manage small messages and real-time
communications (e.g., digital audio and video). It is also one of the highest-bandwidth
networks that Linux currently supports. The drawback is that ATM is not cheap, and there are
still some compatibility problems across vendors. An overview of Linux ATM development is
available at http://lrcwww.epfl.ch/linux-atm/
Design Considerations:
Linux support: kernel driver, AAL* library
Maximum bandwidth: 155 Mb/s (soon, 1,200 Mb/s)
Minimum latency: 120 microseconds
Available as: multiple-vendor hardware
Interface port/bus used: PCI
Network structure: switched hubs
Cost per machine connected: $3,000
Page 92
Cluster Computing
Revision Version 2.3
Copyright © 2003
FC (Fibre Channel)
The goal of FC (Fibre Channel) is to provide high-performance block I/O (an FC frame
carries a 2,048 byte data payload), particularly for sharing disks and other storage devices that
can be directly connected to the FC rather than connected through a computer. Bandwidthwise, FC is specified to be relatively fast, running anywhere between 133 and 1,062 Mbits/s. If
FC becomes popular as a high-end SCSI replacement, it may quickly become a cost-effective
technology; for now, it is not cost-effective and is not supported by Linux. The Fibre Channel
Association at maintains a good collection of FC references:
http://www.amdahl.com/ext/CARP/FCA/FCA.html
Design Considerations:
Linux support: no
Maximum bandwidth: 1,062 Mb/s
Minimum latency: Data not available
Available as: multiple-vendor hardware
Interface port/bus used: PCI?
Network structure: Data not available
Cost per machine connected: Data not available
FireWire (IEEE 1394)
FireWire, http://www.firewire.org/, the IEEE 1394-1995 standard, is destined to be the lowcost high-speed digital network for consumer electronics. The showcase application is
connecting DV digital video camcorders to computers, however FireWire is intended to be
used for applications ranging from being a SCSI replacement to interconnecting the
components of your home theatre. It allows up to 64K devices to be connected in any topology
using busses and bridges that does not create a cycle, and automatically detects the
configuration when components are added or removed. Short (four-byte "quadlet") low-latency
messages are supported as well as ATM-like isochronous transmission (used to keep
multimedia messages synchronized). Adaptec has FireWire products that allow up to 63
devices to be connected to a single PCI interface card, and also has good general FireWire
information at http://www.adaptec.com/serialio/
Although FireWire will not be the highest bandwidth network available, the consumer-level
market (which should drive prices very low) and low latency support might make this one of
the best Linux PC cluster message-passing network technologies within the next year or so.
Design Considerations:
Linux support: No
Maximum bandwidth: 196.608 Mb/s (soon, 393.216 Mb/s)
Minimum latency: Data not available
Available as: multiple-vendor hardware
Interface port/bus used: PCI
Network structure: random without cycles (self-configuring)
Cost per machine connected: $600
Cluster Computing
Page 93
Revision Version 2.3
Copyright © 2003
HiPPI And Serial HiPPI
HiPPI (High Performance Parallel Interface) was originally intended to provide very high
bandwidth for transfer of huge data sets between a supercomputer and another machine (a
supercomputer, frame buffer, disk array, etc.), and has become the dominant standard for
supercomputers. Although it is an oxymoron, Serial HiPPI is also becoming popular, typically
using a fibre optic cable instead of the 32-bit wide standard (parallel) HiPPI cables. Over the
past few years, HiPPI crossbar switches have become common and prices have dropped
sharply; unfortunately, serial HiPPI is still quite expensive, and that is what PCI bus interface
cards generally support. Further to this, Linux doesn't yet support HiPPI. A good overview of
HiPPI is maintained by CERN at http://www.cern.ch/HSI/hippi/; they also maintain a rather
long list of HiPPI vendors at http://www.cern.ch/HSI/hippi/procintf/manufact.htm
Design Considerations:
Linux support: no
Maximum bandwidth: 1,600 Mb/s (serial is 1,200 Mb/s)
Minimum latency: Data not available
Available as: multiple-vendor hardware
Interface port/bus used: EISA, PCI
Network structure: switched hubs
Cost per machine connected: $3,500 (serial is $4,500)
IrDA (Infrared Data Association)
IrDA (Infrared Data Association, http://www.irda.org/) is the infrared device on the side of
many laptop PCs. It is inherently difficult to connect more than two machines using this
interface, so it is highly unlikely to be used for clustering. Don Becker did some preliminary
work with IrDA.
Design Considerations:
Linux support: no1
Maximum bandwidth: 1.15 Mb/s and 4 Mb/s
Minimum latency: Data not available
Available as: multiple-vendor hardware
Interface port/bus used: IrDA
Network structure: None required
Cost per machine connected: $0
Page 94
Cluster Computing
Revision Version 2.3
Copyright © 2003
SCI
The goal of SCI (Scalable Coherent Interconnect, ANSI/IEEE 1596-1992) is to essentially
provide a high performance mechanism that can support coherent shared memory access
across large numbers of machines, as well various types of block message transfers. It is
considered that the designed bandwidth and latency of SCI are both far in advance in
comparison to most other network technologies. The drawback is that SCI is not widely
available as cost-effective production units, and there is not any Linux support.
SCI is used primarily in various proprietary designs for logically shared physically distributed
memory machines, such as the HP/Convex Exemplar SPP and the Sequent NUMA-Q 2000
(see http://www.sequent.com/). However, SCI is available as a PCI interface card and 4-way
switches (up to 16 machines can be connected by cascading four 4-way switches) from
Dolphin, http://www.dolphinics.com/, as their CluStar product line. A good set of links
overviewing SCI is maintained by CERN at http://www.cern.ch/HSI/sci/sci.html
Design Considerations:
Linux support: no
Maximum bandwidth: 4,000 Mb/s
Minimum latency: 2.7 microseconds
Available as: multiple-vendor hardware
Interface port/bus used: PCI, proprietary
Network structure: Data not available
Cost per machine connected: > $1,000
SCSI
SCSI (Small Computer Systems Interconnect) is essentially an I/O bus that is used for disk
drives, CD ROMS, image scanners, and other peripherals. There are three separate standards
SCSI-1, SCSI-2, and SCSI-3; Fast and Ultra speeds; and data path widths of 8, 16, or 32 bits
(with FireWire compatibility also mentioned in SCSI-3). It is quite a complicated set of
standards, however SCSI is somewhat faster than EIDE and can handle more devices more
efficiently.
What many people do not realize is that it is fairly simple for two computers to share a single
SCSI bus. This type of configuration is very useful for sharing disk drives between machines
and implementing fail-over - having one machine take over database requests when the other
machine fails. Currently, this is the only mechanism supported by Microsoft's PC cluster
product, Windows 2000 Cluster Server (formerly known as WolfPack). However, the inability
to scale to larger systems renders shared SCSI limited for parallel processing in general.
Design Considerations:
Linux support: kernel drivers
Maximum bandwidth: 5 Mb/s to over 20 Mb/s
Minimum latency: Data not available
Available as: multiple-vendor hardware
Interface port/bus used: PCI, EISA, ISA card
Network structure: inter-machine bus sharing SCSI devices
Cost per machine connected: Data not available
Cluster Computing
Page 95
Revision Version 2.3
Copyright © 2003
ServerNet
ServerNet is the high-performance network hardware from Compaq (technology obtained from
Tandem). Especially in the online transaction processing (OLTP) world, Compaq is well
known as a leading producer of high-reliability systems, so it is not surprising that their
network claims not just high performance, but also "high data integrity and reliability".
Another interesting aspect of ServerNet is that it claims to be able to transfer data from any
device directly to any device; not just between processors, but also disk drives, etc., in a onesided style similar to that suggested by the MPI remote memory access mechanisms. While a
single vendor only offers ServerNet, that vendor is powerful enough to potentially establish
ServerNet as a major standard.
Design Considerations:
Linux support: no
Maximum bandwidth: 400 Mb/s
Minimum latency: 3 microseconds
Available as: single-vendor hardware
Interface port/bus used: PCI
Network structure: hexagonal tree/tetrahedral lattice of hubs
Cost per machine connected: Data not available
SHRIMP
The SHRIMP project, http://www.CS.Princeton.EDU/shrimp/, at the Princeton University
Computer Science Department is building a parallel computer using PCs running Linux as the
processing elements. The first SHRIMP (Scalable, High-Performance, Really Inexpensive
Multi-Processor) was a simple two-processor prototype using a dual-ported RAM on a custom
EISA card interface. There is now a prototype that will scale to larger configurations using a
custom interface card to connect to a "hub" that is essentially the same mesh routing network
used in the Intel Paragon (see http://www.ssd.intel.com/paragon.html). Considerable effort has
gone into developing low-overhead "virtual memory mapped communication" hardware and
support software.
Design Considerations:
Linux support: user-level memory mapped interface
Maximum bandwidth: 180 Mb/s
Minimum latency: 5 microseconds
Available as: research prototype
Interface port/bus used: EISA
Network structure: mesh backplane (as in Intel Paragon)
Cost per machine connected: Data not available
Page 96
Cluster Computing
Revision Version 2.3
Copyright © 2003
TTL_PAPERS
The PAPERS (Purdue's Adapter for Parallel Execution and Rapid Synchronization) project,
http://garage.ecn.purdue.edu/~papers/, at the Purdue University School of Electrical and
Computer Engineering is building scalable, low-latency, aggregate function communication
hardware and software that allows a parallel supercomputer to be built using unmodified
PCs/workstations as nodes.
There have been over a dozen different types of PAPERS hardware built that connect to
PCs/workstations via the SPP (Standard Parallel Port), roughly following two development
lines. The versions called "PAPERS" target higher performance, using whatever technologies
are appropriate; current work uses FPGAs, and high bandwidth PCI bus interface designs are
also under development. In contrast, the versions called "TTL_PAPERS" are designed to be
easily reproduced outside Purdue, and are remarkably simple public domain designs that can
be built using ordinary TTL logic. One such design is produced commercially,
http://chelsea.ios.com/~hgdietz/sbm4.html
Unlike the custom hardware designs from other universities, TTL_PAPERS clusters have been
assembled at many universities from the USA to South Korea. Bandwidth is severely limited
by the SPP connections, but PAPERS implements very low latency aggregate function
communications; even the fastest message-oriented systems cannot provide comparable
performance on those aggregate functions. Thus, PAPERS is particularly good for
synchronizing the displays of a video wall, scheduling accesses to a high-bandwidth network,
evaluating global fitness in genetic searches, etc. Although PAPERS clusters have been built
using IBM PowerPC AIX, DEC Alpha OSF/1, and HP PA-RISC HP-UX machines, Linuxbased PCs are the platforms best supported.
User programs using TTL_PAPERS AFAPI directly access the SPP hardware port registers
under Linux, without an OS call for each access. To do this, AFAPI first gets port permission
using either iopl() or ioperm(). The problem with these calls is that both require the user
program to be privileged, yielding a potential security hole. The solution is an optional kernel
patch, http://garage.ecn.purdue.edu/~papers/giveioperm.html that allows a privileged process
to control port permission for any process.
Design Considerations:
Linux support: AFAPI library
Maximum bandwidth: 1.6 Mb/s
Minimum latency: 3 microseconds
Available as: public-domain design, single-vendor hardware
Interface port/bus used: SPP
Network structure: tree of hubs
Cost per machine connected: $100
Cluster Computing
Page 97
Revision Version 2.3
Copyright © 2003
WAPERS
WAPERS (Wired-AND Adapter for Parallel Execution and Rapid Synchronization) is a spin-off of
the PAPERS project, http://garage.ecn.purdue.edu/~papers/, at the Purdue University School of
Electrical and Computer Engineering. If implemented properly, the SPP has four bits of opencollector output that can be wired together across machines to implement a 4-bit wide wired
AND. This wired-AND is electrically touchy, and the maximum number of machines that can
be connected in this way critically depends on the analog properties of the ports (maximum
sink current and pull-up resistor value); typically, up to 7 or 8 machines can be networked by
WAPERS. Although cost and latency are very low, so is bandwidth; WAPERS is much better
as a second network for aggregate operations than as the only network in a cluster. As with
TTL_PAPERS, to improve system security, there is a minor kernel patch recommended, but
not required: http://garage.ecn.purdue.edu/~papers/giveioperm.html
Design Considerations:
Linux support: AFAPI library
Maximum bandwidth: 0.4 Mb/s
Minimum latency: 3 microseconds
Available as: public-domain design
Interface port/bus used: SPP
Network structure: wiring pattern between 2-64 machines
Cost per machine connected: $5
Page 98
Cluster Computing
Revision Version 2.3
Copyright © 2003
8.2. Appendix B – Channel Bonding
/* Mode: C;
* ifenslave.c: Configure network interfaces for parallel routing.
*
*
This program controls the Linux implementation of running multiple
*
network interfaces in parallel.
*
* Usage:
ifenslave [-v] master-interface slave-interface [metric <N>]
*
* Author:
Donald Becker <[email protected]>
*
Copyright 1994-1996 Donald Becker
*
*
This program is free software; you can redistribute it
*
and/or modify it under the terms of the GNU General Public
*
License as published by the Free Software Foundation.
*
*
The author may be reached as [email protected], or C/O
*
Center of Excellence in Space Data and Information Sciences
*
Code 930.5, Goddard Space Flight Center, Greenbelt MD 20771
*/
static char *version =
"ifenslave.c:v0.07 9/9/97 Donald Becker ([email protected])\n";
static const char *usage_msg =
"Usage: ifenslave [-afrvV] <master-interface> <slave-interface> [metric <N>]\n";
#include
#include
#include
#include
#include
#include
#include
#include
#include
#include
#include
#include
#include
#include
<unistd.h>
<stdlib.h>
<stdio.h>
<ctype.h>
<string.h>
<errno.h>
<fcntl.h>
<getopt.h>
<sys/types.h>
<sys/socket.h>
<sys/ioctl.h>
<linux/if.h>
<linux/if_arp.h>
<linux/if_ether.h>
struct option longopts[] = {
/* { name has_arg *flag val } */
{"all-interfaces", 0, 0, 'a'},
/* Show all interfaces. */
{"force",
0, 0, 'f'},
/* Force the operation. */
{"help",
0, 0, '?'},
/* Give help */
{"receive-slave", 0, 0, 'r'},
/* Make a receive-only slave. */
{"verbose",
0, 0, 'v'},
/* Report each action taken.
{"version",
0, 0, 'V'},
/* Emit version information.
{ 0, 0, 0, 0 }
};
/* Command-line flags. */
unsigned int
opt_a = 0,
opt_f = 0,
opt_r = 0,
verbose = 0,
opt_version;
int skfd = -1;
/*
/*
/*
/*
*/
*/
Show-all-interfaces flag. */
Force the operation. */
Set up a Rx-only slave. */
Verbose flag. */
/* AF_INET socket for ioctl() calls. */
static void if_print(char *ifname);
int
main(int argc, char **argv)
{
struct ifreq ifr2, if_hwaddr, if_ipaddr, if_metric, if_mtu, if_dstaddr;
struct ifreq if_netmask, if_brdaddr, if_flags;
int goterr = 0;
int c, errflag = 0;
char **spp, *master_ifname, *slave_ifname;
while ((c = getopt_long(argc, argv, "afrvV?", longopts, 0)) != EOF)
switch (c) {
case 'a': opt_a++; break;
case 'f': opt_f++; break;
case 'r': opt_r++; break;
Cluster Computing
Page 99
Revision Version 2.3
Copyright © 2003
case 'v': verbose++;
case 'V': opt_version++;
case '?': errflag++;
}
if (errflag) {
fprintf(stderr, usage_msg);
return 2;
}
break;
break;
if (verbose || opt_version)
printf(version);
/* Open a basic socket. */
if ((skfd = socket(AF_INET, SOCK_DGRAM,0)) < 0) {
perror("socket");
exit(-1);
}
#ifdef notdef
/* Find options scattered throughout the command line.
I should change this to use getopt() sometime. */
argc--; argv++;
while (*argv[0] == '-') {
char *argp = *argv++;
argc--;
while (*++argp) {
switch (*argp) {
default:
fprintf(stderr, "Unrecognized option '%c'.\n%s",
argp[0], usage_msg);
return 2;
}
}
}
#endif
if (verbose)
fprintf(stderr, "DEBUG: argc=%d, optind=%d and argv[optind] is %s.\n",
argc, optind, argv[optind]);
/* No remaining args means show all interfaces. */
if (optind == argc) {
if_print((char *)NULL);
(void) close(skfd);
exit(0);
}
/* Copy the interface name. */
spp = argv + optind;
master_ifname = *spp++;
slave_ifname = *spp++;
/* A single args means show the configuration for this interface. */
if (slave_ifname == NULL) {
if_print(master_ifname);
(void) close(skfd);
exit(0);
}
/* Get the vitals from the master interface. */
strncpy(if_hwaddr.ifr_name, master_ifname, IFNAMSIZ);
if (ioctl(skfd, SIOCGIFHWADDR, &if_hwaddr) < 0) {
fprintf(stderr, "SIOCGIFHWADDR on %s failed: %s\n", master_ifname,
strerror(errno));
return 1;
} else {
/* Gotta convert from 'char' to unsigned for printf(). */
unsigned char *hwaddr = (unsigned char *)if_hwaddr.ifr_hwaddr.sa_data;
/* The family '1' is ARPHRD_ETHER for ethernet. */
if (if_hwaddr.ifr_hwaddr.sa_family != 1 && !opt_f) {
fprintf(stderr, "The specified master interface '%s' is not"
" ethernet-like.\n This program is designed to work"
" with ethernet-like network interfaces.\n"
" Use the '-f' option to force the operation.\n",
master_ifname);
return 1;
}
if (verbose)
printf("The hardware address (SIOCGIFHWADDR) of %s is type %d "
"%2.2x:%2.2x:%2.2x:%2.2x:%2.2x:%2.2x.\n", master_ifname,
if_hwaddr.ifr_hwaddr.sa_family, hwaddr[0], hwaddr[1],
Page 100
Cluster Computing
Revision Version 2.3
Copyright © 2003
hwaddr[2], hwaddr[3], hwaddr[4], hwaddr[5]);
}
{
struct ifreq *ifra[6] = { &if_ipaddr, &if_mtu, &if_dstaddr,
&if_brdaddr, &if_netmask,
&if_flags };
const char *req_name[6] = {
"IP address", "MTU", "destination address",
"broadcast address", "netmask", "status flags", };
const int ioctl_req_type[6] = {
SIOCGIFADDR, SIOCGIFMTU, SIOCGIFDSTADDR,
SIOCGIFBRDADDR, SIOCGIFNETMASK, SIOCGIFFLAGS, };
int i;
for (i = 0; i < 6; i++) {
strncpy(ifra[i]->ifr_name, master_ifname, IFNAMSIZ);
if (ioctl(skfd, ioctl_req_type[i], ifra[i]) < 0) {
fprintf(stderr,
"Something broke getting the master's %s: %s.\n",
req_name[i], strerror(errno));
}
}
}
do {
strncpy(ifr2.ifr_name, slave_ifname, IFNAMSIZ);
if (ioctl(skfd, SIOCGIFFLAGS, &ifr2) < 0) {
int saved_errno = errno;
fprintf(stderr, "SIOCGIFFLAGS on %s failed: %s\n", slave_ifname,
strerror(saved_errno));
return 1;
}
if (ifr2.ifr_flags & IFF_UP) {
printf("The interface %s is up, shutting it down it to enslave it.\n",
slave_ifname);
ifr2.ifr_flags &= ~IFF_UP;
if (ioctl(skfd, SIOCSIFFLAGS, &ifr2) < 0) {
int saved_errno = errno;
fprintf(stderr, "Shutting down interface %s failed: %s\n",
slave_ifname, strerror(saved_errno));
}
}
strncpy(if_hwaddr.ifr_name, slave_ifname, IFNAMSIZ);
if (ioctl(skfd, SIOCSIFHWADDR, &if_hwaddr) < 0) {
int saved_errno = errno;
fprintf(stderr, "SIOCSIFHWADDR on %s failed: %s\n", if_hwaddr.ifr_name,
strerror(saved_errno));
if (saved_errno == EBUSY)
fprintf(stderr, " The slave device %s is busy: it must be"
" idle before running this command.\n",
slave_ifname);
else if (saved_errno == EOPNOTSUPP)
fprintf(stderr, " The slave device you specified does not
support"
" setting the MAC address.\n
Your kernel likely
does not"
" support slave devices.\n");
else if (saved_errno == EINVAL)
fprintf(stderr, " The slave device's address type does not
match"
" the master's address type.\n");
return 1;
} else {
if (verbose) {
unsigned char *hwaddr = if_hwaddr.ifr_hwaddr.sa_data;
printf("Set the slave's hardware address to "
"%2.2x:%2.2x:%2.2x:%2.2x:%2.2x:%2.2x.\n", hwaddr[0],
hwaddr[1], hwaddr[2], hwaddr[3], hwaddr[4],
hwaddr[5]);
}
}
if (*spp && !strcmp(*spp, "metric")) {
if (*++spp == NULL) {
fprintf(stderr, usage_msg);
exit(2);
}
if_metric.ifr_metric = atoi(*spp);
if (ioctl(skfd, SIOCSIFMETRIC, &if_metric) < 0) {
fprintf(stderr, "SIOCSIFMETRIC: %s\n", strerror(errno));
goterr = 1;
Cluster Computing
Page 101
Revision Version 2.3
Copyright © 2003
}
spp++;
}
if (strncpy(if_ipaddr.ifr_name, slave_ifname, IFNAMSIZ) <= 0
|| ioctl(skfd, SIOCSIFADDR, &if_ipaddr) < 0) {
fprintf(stderr,
"Something broke setting the slave's address: %s.\n",
strerror(errno));
} else {
if (verbose) {
unsigned char *ipaddr = if_ipaddr.ifr_addr.sa_data;
printf("Set the slave's IP address to %d.%d.%d.%d.\n",
ipaddr[0], ipaddr[1], ipaddr[2], ipaddr[3]);
}
}
if (strncpy(if_mtu.ifr_name, slave_ifname, IFNAMSIZ) <= 0
|| ioctl(skfd, SIOCSIFMTU, &if_mtu) < 0) {
fprintf(stderr, "Something broke setting the slave MTU: %s.\n",
strerror(errno));
} else {
if (verbose)
printf("Set the slave's MTU to %d.\n", if_mtu.ifr_mtu);
}
if (strncpy(if_dstaddr.ifr_name, slave_ifname, IFNAMSIZ) <= 0
|| ioctl(skfd, SIOCSIFDSTADDR, &if_dstaddr) < 0) {
fprintf(stderr, "Error setting the slave with SIOCSIFDSTADDR: %s.\n",
strerror(errno));
} else {
if (verbose) {
unsigned char *ipaddr = if_dstaddr.ifr_dstaddr.sa_data;
printf("Set the slave's destination address to %d.%d.%d.%d.\n",
ipaddr[0], ipaddr[1], ipaddr[2], ipaddr[3]);
}
}
if (strncpy(if_brdaddr.ifr_name, slave_ifname, IFNAMSIZ) <= 0
|| ioctl(skfd, SIOCSIFBRDADDR, &if_brdaddr) < 0) {
fprintf(stderr,
"Something broke setting the slave broadcast address:
%s.\n",
strerror(errno));
} else {
if (verbose) {
unsigned char *ipaddr = if_brdaddr.ifr_broadaddr.sa_data;
printf("Set the slave's broadcast address to %d.%d.%d.%d.\n",
ipaddr[0], ipaddr[1], ipaddr[2], ipaddr[3]);
}
}
if (strncpy(if_netmask.ifr_name, slave_ifname, IFNAMSIZ) <= 0
|| ioctl(skfd, SIOCSIFNETMASK, &if_netmask) < 0) {
fprintf(stderr,
"Something broke setting the slave netmask: %s.\n",
strerror(errno));
} else {
if (verbose) {
unsigned char *ipaddr = if_netmask.ifr_netmask.sa_data;
printf("Set the slave's netmask to %d.%d.%d.%d.\n",
ipaddr[0], ipaddr[1], ipaddr[2], ipaddr[3]);
}
}
if ((if_flags.ifr_flags &= ~(IFF_SLAVE | IFF_MASTER)) == 0
|| strncpy(if_flags.ifr_name, slave_ifname, IFNAMSIZ) <= 0
|| ioctl(skfd, SIOCSIFFLAGS, &if_flags) < 0) {
fprintf(stderr,
"Something broke setting the slave flags: %s.\n",
strerror(errno));
} else {
if (verbose)
printf("Set the slave's flags %4.4x.\n", if_flags.ifr_flags);
}
/* Do the real thing: set the second interface as a slave. */
if ( ! opt_r) {
strncpy(if_flags.ifr_name, master_ifname, IFNAMSIZ);
strncpy(if_flags.ifr_slave, slave_ifname, IFNAMSIZ);
if (ioctl(skfd, SIOCSIFSLAVE, &if_flags) < 0) {
fprintf(stderr,
"SIOCSIFSLAVE: %s.\n", strerror(errno));
}
Page 102
Cluster Computing
Revision Version 2.3
Copyright © 2003
}
} while ( (slave_ifname = *spp++) != NULL);
/* Close the socket. */
(void) close(skfd);
return(goterr);
}
static short mif_flags;
/* Get the inteface configuration from the kernel. */
static int if_getconfig(char *ifname)
{
struct ifreq ifr;
int metric, mtu;
/* Parameters of the master interface. */
struct sockaddr dstaddr, broadaddr, netmask;
strcpy(ifr.ifr_name, ifname);
if (ioctl(skfd, SIOCGIFFLAGS, &ifr) < 0)
return -1;
mif_flags = ifr.ifr_flags;
printf("The result of SIOCGIFFLAGS on %s is %x.\n",
ifname, ifr.ifr_flags);
strcpy(ifr.ifr_name, ifname);
if (ioctl(skfd, SIOCGIFADDR, &ifr) < 0)
return -1;
printf("The result of SIOCGIFADDR is %2.2x.%2.2x.%2.2x.%2.2x.\n",
ifr.ifr_addr.sa_data[0], ifr.ifr_addr.sa_data[1],
ifr.ifr_addr.sa_data[2], ifr.ifr_addr.sa_data[3]);
strcpy(ifr.ifr_name, ifname);
if (ioctl(skfd, SIOCGIFHWADDR, &ifr) < 0)
return -1;
{
/* Gotta convert from 'char' to unsigned for printf(). */
unsigned char *hwaddr = (unsigned char *)ifr.ifr_hwaddr.sa_data;
printf("The result of SIOCGIFHWADDR is type %d "
"%2.2x:%2.2x:%2.2x:%2.2x:%2.2x:%2.2x.\n",
ifr.ifr_hwaddr.sa_family, hwaddr[0], hwaddr[1],
hwaddr[2], hwaddr[3], hwaddr[4], hwaddr[5]);
}
strcpy(ifr.ifr_name, ifname);
if (ioctl(skfd, SIOCGIFMETRIC, &ifr) < 0) {
metric = 0;
} else
metric = ifr.ifr_metric;
strcpy(ifr.ifr_name, ifname);
if (ioctl(skfd, SIOCGIFMTU, &ifr) < 0)
mtu = 0;
else
mtu = ifr.ifr_mtu;
strcpy(ifr.ifr_name, ifname);
if (ioctl(skfd, SIOCGIFDSTADDR, &ifr) < 0) {
memset(&dstaddr, 0, sizeof(struct sockaddr));
} else
dstaddr = ifr.ifr_dstaddr;
strcpy(ifr.ifr_name, ifname);
if (ioctl(skfd, SIOCGIFBRDADDR, &ifr) < 0) {
memset(&broadaddr, 0, sizeof(struct sockaddr));
} else
broadaddr = ifr.ifr_broadaddr;
strcpy(ifr.ifr_name, ifname);
if (ioctl(skfd, SIOCGIFNETMASK, &ifr) < 0) {
memset(&netmask, 0, sizeof(struct sockaddr));
} else
netmask = ifr.ifr_netmask;
return(0);
}
static void if_print(char *ifname)
{
char buff[1024];
struct ifconf ifc;
struct ifreq *ifr;
int i;
Cluster Computing
Page 103
Revision Version 2.3
Copyright © 2003
if (ifname == (char *)NULL) {
ifc.ifc_len = sizeof(buff);
ifc.ifc_buf = buff;
if (ioctl(skfd, SIOCGIFCONF, &ifc) < 0) {
fprintf(stderr, "SIOCGIFCONF: %s\n", strerror(errno));
return;
}
ifr = ifc.ifc_req;
for (i = ifc.ifc_len / sizeof(struct ifreq); --i >= 0; ifr++) {
if (if_getconfig(ifr->ifr_name) < 0) {
fprintf(stderr, "%s: unknown interface.\n",
ifr->ifr_name);
continue;
}
if (((mif_flags & IFF_UP) == 0) && !opt_a) continue;
/*ife_print(&ife);*/
}
} else {
if (if_getconfig(ifname) < 0)
fprintf(stderr, "%s: unknown interface.\n", ifname);
}
}
_
/*
* Local variables:
* version-control: t
* kept-new-versions: 5
* c-indent-level: 4
* c-basic-offset: 4
* tab-width: 4
* compile-command: "gcc -Wall -Wstrict-prototypes -O ifenslave.c -o ifenslave"
* End:
*/
Page 104
Cluster Computing
Revision Version 2.3
Copyright © 2003
8.3. Appendix C – MPI Implementations
The following MPI Implementations were considered for use [19].
Cluster Computing
Page 105
Revision Version 2.3
8.3.1.
Copyright © 2003
LAM
LAM/MPI
University of Notre Dame
6.5.1
05/06/2001 at 08:58:22
http://www.lam-mpi.org/
[email protected]
Just about all POSIX platforms.
Yes
No
Yes
Topic
Supplier:
Version:
Last update:
URL:
E-mail:
Platforms:
Source code available:
Commerical:
Free:
All (excluding MPI_CANCEL)
MPI-1:
MPI_CANCEL
mpiexec
MPI_INIT(NULL, NULL)
Supported
TYPE_INDEXED_BLOCK
No Support
STATUS_IGNORE
Supported
Keyval error class
Supported
Committed datatype
Supported
User functions @ termination
No Support
MPI_FINALIZED
Supported
MPI_Info
Supported
MPI_MALLOC
Supported
Language interoperability
Supported
Error handlers
Supported
New Datatypes
Canonical [UN]PACK
Supported
Establishing communication
Supported
Other
Supported
Synchronization
Some Support
Error handling
Some Support
Intercommunicator constructors
Extended collectives
Availability
Notes
Status information
No Support
Naming objects
Supported
Error classes
No Support
Threads
Supported
Duplicating datatypes
Topic
Notes
MPI_INIT_THREAD only supports
MPI THREAD SINGLE
Supported
Availability
File manipulation
Supported
File views
Supported
Data access
Notes
Supported
Some Support
Attribute caching
Missing MPI_Win_test, MPI_Win_lock,
and MPI Win unlock
Have MPI_ERR_WIN and
MPI ERR BASE
No Support
Availability
No Support
Notes
Supported
File interoperability
Some Support
Consistency
Supported
Error handling
No Support
Error classes
No Support
Topic
Availability
C++ for MPI-1.2
Supported
C++ for MPI-2
No Support
Fortran 90 module
Topic
IMPI support
Notes
No Support
Generalized requests
Decoding datatypes
IMPI
Availability
Communication
Topic
MPI-2: Language
Have MPI_WCHAR and
MPI_UNSIGNED_LONG_LONG; missing
MPI SIGNED CHAR
Supported
Topic
MPI-2: I/O
Notes
Supported
Initialization
MPI-2: External Interfaces
No Support
Availability
Process manager
Topic
MPI-2: Extended Collective
Operation
Canceling sends is not supported
Supported
Some Support
Topic
MPI-2: One-sided
Communication
Availability
No Support
Datatype manipulations
MPI-2: Process Creation and
Management
Notes
Supported
Some Support
Topic
MPI-2: Miscellany
Availability
Notes
No Support
Availability
Some Support
Notes
Point to point functionality supported
Table 8-1 – LAM MPI 2 Implementation Details
Page 106
Cluster Computing
Revision Version 2.3
8.3.2.
Copyright © 2003
MPICH
MPICH
Argonne National Laboratory
1.2.1
11/13/2000 at 08:17:26
http://www-unix.mcs.anl.gov/mpi/mpich/
[email protected]
MPP's, IBM SP, Intel Paragon, SGI Onyx, Challenge and Power Challenge, Convex (HP)
Yes
No
Yes
Topic
Availability
Notes
Supplier:
Version:
Last update:
URL:
E-mail:
Platforms:
Source code available:
Commerical:
Free:
All (excluding MPI_CANCEL)
MPI-1:
MPI_CANCEL
Supported
Topic
MPI-2: Miscellany
No Support
MPI_INIT(NULL, NULL)
No Support
TYPE_INDEXED_BLOCK
Supported
STATUS_IGNORE
No Support
Keyval error class
Supported
Committed datatype
Supported
User functions @ termination
No Support
MPI_FINALIZED
Supported
MPI_Info
Supported
MPI_MALLOC
No Support
Language interoperability
Supported
Error handlers
No Support
Datatype manipulations
Some Support
New Datatypes
Some Support
Topic
No Support
Establishing communication
No Support
Other
No Support
Communication
No Support
Synchronization
No Support
Notes
No Support
Intercommunicator constructors
Extended collectives
Topic
Generalized requests
Status information
Naming objects
Availability
Notes
No Support
No Support
Availability
Notes
No Support
Supported
Some Support
Error classes
No Support
Decoding datatypes
Threads
Supported
Some Support
Attribute caching
No Support
Duplicating datatypes
Topic
No Support
Availability
File manipulation
Supported
File views
Supported
Data access
Notes
Supported
File interoperability
MPI-2: I/O
Availability
Initialization
Topic
MPI-2: External Interfaces
Notes
No Support
Error handling
MPI-2: Extended Collective
Operation
Notes
No Support
Availability
Process manager
Topic
MPI-2: One-sided
Communication
Availability
mpiexec
Canonical [UN]PACK
MPI-2: Process Creation and
Management
Supported
Some Support
Consistency
Supported
Error handling
No Support
Error classes
No Support
Topic
C++ for MPI-1.2
Availability
Notes
Supported
C++ for MPI-2
No Support
MPI-2: Language
Fortran 90 module
IMPI
IMPI support
Topic
Some Support
Availability
Notes
No Support
Table 8-2 – MPICH MPI 2 Implementation Details
Cluster Computing
Page 107
Revision Version 2.3
Copyright © 2003
8.4. Appendix D – POV-Ray
The following appendix contains some of the POV-Ray test script files or .pov files used on
the test cluster as well as performance data extracted from the dataset on comparable clusters.
[43.2]
8.4.1.
POV-Ray Benchmark
Test Settings
For command line versions:
povray -i skyvase.pov +v1 -d +ft -x +a0.300 +r3 -q9 -w640 -h480 -mv2.0 +b1000 > results.txt
(For Unix/Linux versions prepend with "time" and note the CPU time used)
For GUI versions:
Go to menu: Render --> Edit settings/Render (or press Alt+C) and paste:
+v1 -d +ft -x +a0.300 +r3 -q9 -w640 -h480 -mv2.0 +b1000
into the "Command line options" box.
This will temporarily override your render settings.
If you're running a multi-tasking OS, try to give POV full attention.
skyvase.pov file
// Persistence Of Vision raytracer version 2.0 sample file.
// By Dan Farmer
//
Minneapolis, mn
//
skyvase.pov
//
Vase made with Hyperboloid and sphere {, sitting on a hexagonal
//
marble column. Take note of the color and surface characteristics
//
of the gold band around the vase. It seems to be a successful
//
combination for gold or brass.
//
// Contains a Disk_Y object which may have changed in shapes.dat
#include
#include
#include
#include
"shapes.inc"
"shapes2.inc"
"colors.inc"
"textures.inc"
#declare DMF_Hyperboloid = quadric {
<1.0, -1.0, 1.0>,
<0.0, 0.0, 0.0>,
<0.0, 0.0, 0.0>,
-0.5
}
/* Like Hyperboloid_Y, but more curvy */
camera {
location <0.0, 28.0, -200.0>
direction <0.0, 0.0, 2.0>
up <0.0, 1.0, 0.0>
right <4/3, 0.0, 0.0>
look_at <0.0, -12.0, 0.0>
}
/* Light behind viewer postion (pseudo-ambient light) */
light_source { <100.0, 500.0, -500.0> colour White }
union {
union {
intersection {
plane { y, 0.7 }
Page 108
Cluster Computing
Revision Version 2.3
Copyright © 2003
object { DMF_Hyperboloid scale <0.75, 1.25, 0.75> }
object { DMF_Hyperboloid scale <0.70, 1.25, 0.70> inverse }
plane { y, -1.0 inverse }
}
sphere { <0, 0, 0>, 1 scale <1.6, 0.75, 1.6 > translate <0, -1.15, 0> }
scale <20, 25, 20>
pigment {
Bright_Blue_Sky
turbulence 0.3
quick_color Blue
scale <8.0, 4.0, 4.0>
rotate 15*z
}
finish {
ambient 0.1
diffuse 0.75
phong 1
phong_size 100
reflection 0.35
}
}
sphere { /* Gold ridge around sphere portion of vase*/
<0, 0, 0>, 1
scale <1.6, 0.75, 1.6>
translate -7*y
scale <20.5, 4.0, 20.5>
finish { Metal }
pigment { OldGold }
}
bounded_by {
object {
Disk_Y
translate -0.5*y // Remove for new Disk_Y definition
scale <34, 100, 34>
}
}
}
/* Stand for the vase */
object { Hexagon
rotate -90.0*z
rotate -45*y
scale <40, 25, 40>
translate -70*y
/* Stand it on end (vertical)*/
/* Turn it to a pleasing angle */
pigment {
Sapphire_Agate
quick_color Red
scale 10.0
}
finish {
ambient 0.2
diffuse 0.75
reflection 0.85
}
}
union {
plane { z, 50
plane { z, 50
rotate -45*y }
rotate +45*y }
pigment { DimGray }
finish {
ambient 0.2
diffuse 0.75
reflection 0.5
}
}
Cluster Computing
Page 109
Revision Version 2.3
8.4.2.
Copyright © 2003
Alternate Bench Mark POV Files
poolballs.pov
/*
DEMONSTRATION FILE
POOLBALL PLUGIN (version 1.0)
(C) Nathan G B O'Brien 1997
[email protected]
http://www.ozemail.com.au/~no13
*/
// ==== Standard POV-Ray Includes ====
#include "colors.inc" // Standard Color definitions
#include "textures.inc"
// Standard Texture definitions
camera {
location <0,20,-30>
look_at <0,10,0>
}
light_source {<10,50,-20> color White}
// Step one.
// Declare variables that will be the same for all poolballs
// POOLBALL INFORMATION
#declare Poolball_radius
#declare Poolball_font =
#declare Poolball_insert
#declare Poolball_finish
= 10
"crystal.ttf"
= 1
= finish {phong 1 reflection 0.25}
// FONT ADJUSTMENTS
#declare Small_numberx
#declare Small_numbery
#declare Large_numberx
#declare Large_numbery
0
0.1
0.05
0.1
=
=
=
=
// TABLE SURFACE
box {<500,0,500><-500,-10,-500>
pigment{Green}
finish {phong 1}
normal {bumps .2 scale .5}
}
// EIGHT BALL
#declare Poolball_texture = texture{pigment{red 0 green 0 blue 0}}
#declare Poolball_number = 8
#declare Poolball_band = 0
#declare Pool_rotatex = 0
#declare Pool_rotatey = 0
#declare Pool_rotatez = 0
#include "13_Pball.inc"
object {Poolball translate <0,0,10>}
// THIRTEEN BALL
#declare Poolball_texture = texture{pigment{red 1 green 1 blue 0}}
#declare Poolball_number = 13
#declare Poolball_band = 1
#declare Pool_rotatex = 10
#declare Pool_rotatey = -20
#declare Pool_rotatez = 10
#include "13_Pball.inc"
object {Poolball translate <-15,0,0>}
// FOURTEEN BALL
#declare Poolball_texture = texture{pigment{red 0 green 0 blue 1}}
Page 110
Cluster Computing
Revision Version 2.3
Copyright © 2003
#declare Poolball_number = 14
#declare Poolball_band = 1
#declare Pool_rotatex = 20
#declare Pool_rotatey = 35
#declare Pool_rotatez = -10
#include "13_Pball.inc"
object {Poolball translate <15,0,15>}
// SIX BALL
#declare Poolball_texture = texture{pigment{red 1 green 0 blue 0}}
#declare Poolball_number = 6
#declare Poolball_band = 0
#declare Pool_rotatex = 0
#declare Pool_rotatey = 60
#declare Pool_rotatez = 0
#include "13_Pball.inc"
object {Poolball translate <15,0,-10>}
Poolballs.inc
/*
DEMONSTRATION FILE
POOLBALL PLUGIN (version 1.0)
(C) Nathan G B O'Brien 1997
[email protected]
http://www.ozemail.com.au/~no13
*/
// ==== Standard POV-Ray Includes ====
#include "colors.inc" // Standard Color definitions
#include "textures.inc"
// Standard Texture definitions
camera {
location <0,20,-30>
look_at <0,10,0>
}
light_source {<10,50,-20> color White}
// Step one.
// Declare variables that will be the same for all poolballs
// POOLBALL INFORMATION
#declare Poolball_radius
#declare Poolball_font =
#declare Poolball_insert
#declare Poolball_finish
= 10
"crystal.ttf"
= 1
= finish {phong 1 reflection 0.25}
// FONT ADJUSTMENTS
#declare Small_numberx
#declare Small_numbery
#declare Large_numberx
#declare Large_numbery
0
0.1
0.05
0.1
=
=
=
=
// TABLE SURFACE
box {<500,0,500><-500,-10,-500>
pigment{Green}
finish {phong 1}
normal {bumps .2 scale .5}
}
// EIGHT BALL
#declare Poolball_texture = texture{pigment{red 0 green 0 blue 0}}
#declare Poolball_number = 8
#declare Poolball_band = 0
#declare Pool_rotatex = 0
#declare Pool_rotatey = 0
#declare Pool_rotatez = 0
#include "13_Pball.inc"
object {Poolball translate <0,0,10>}
Cluster Computing
Page 111
Revision Version 2.3
Copyright © 2003
// THIRTEEN BALL
#declare Poolball_texture = texture{pigment{red 1 green 1 blue 0}}
#declare Poolball_number = 13
#declare Poolball_band = 1
#declare Pool_rotatex = 10
#declare Pool_rotatey = -20
#declare Pool_rotatez = 10
#include "13_Pball.inc"
object {Poolball translate <-15,0,0>}
// FOURTEEN BALL
#declare Poolball_texture = texture{pigment{red 0 green 0 blue 1}}
#declare Poolball_number = 14
#declare Poolball_band = 1
#declare Pool_rotatex = 20
#declare Pool_rotatey = 35
#declare Pool_rotatez = -10
#include "13_Pball.inc"
object {Poolball translate <15,0,15>}
// SIX BALL
#declare Poolball_texture = texture{pigment{red 1 green 0 blue 0}}
#declare Poolball_number = 6
#declare Poolball_band = 0
#declare Pool_rotatex = 0
#declare Pool_rotatey = 60
#declare Pool_rotatez = 0
#include "13_Pball.inc"
object {Poolball translate <15,0,-10>}
Page 112
Cluster Computing
Revision Version 2.3
Copyright © 2003
tulips.pov
#declare TulipHeadTexture = texture
{
pigment
{
bumps
color_map {
[0.0 colour rgb<0.9, 0.0, 0.0>]
[1.0 colour rgb<0.8, 0.0, 0.2>]
}
scale <0.005, 0.02, 0.005>
}
normal
{
bumps 0.25
scale <0.025, 1, 0.025>
}
finish
{
phong 0.5
phong_size 25
}
}
#declare TulipStemTexture = texture
{
pigment
{
bumps
color_map {
[0.0 colour rgb<0.5, 0.5, 0.4>]
[1.0 colour rgb<0.3, 0.4, 0.1>]
}
scale <0.005, 0.08, 0.005>
}
normal
{
bumps 0.25
scale <0.025, 1, 0.025>
}
finish
{
ambient 0.0
}
}
#declare GrassTexture = texture
{
pigment
{
bumps
color_map {
[0.0 colour rgb<0.6, 0.6, 0.6>]
[1.0 colour rgb<0.3, 0.7, 0.2>]
}
scale <0.005, 0.5, 0.005>
}
normal
{
bumps 0.25
Cluster Computing
Page 113
Revision Version 2.3
Copyright © 2003
scale <0.03, 2, 0.03>
}
finish
{
ambient 0.0
diffuse 0.5
}
}
#declare WindmillBladesTexture = texture
{
pigment
{
colour rgb<0.2, 0.2, 0.0>
}
}
#declare WindmillBaseTexture = texture
{
pigment
{
bumps
color_map {
[0.0 colour rgb<0.5, 0.5, 0.4>]
[1.0 colour rgb<0.8, 0.8, 0.8>]
}
scale 2
}
}
#declare PetalBlob = blob
{
threshold 0.5
sphere{<0.00, 0, 0> 1.0, 1}
sphere{<0.03, 0, 0> 0.6, -1}
bounded_by{box{<-0.55,-0.4,-0.4>, <-0.3, 0.4, 0.4>}}
}
#declare Petal = union
{
difference
{
object
{
PetalBlob
scale <1, 2, 1>
}
plane
{
<0, 1, 0>, 0
}
}
difference
{
object
{
PetalBlob
scale <1, 1, 1>
}
plane
{
<0, -1, 0>, 0
}
}
Page 114
Cluster Computing
Revision Version 2.3
Copyright © 2003
rotate <0, 0, 10>
translate <0.05, 0, 0>
}
#declare TulipHead = union
{
#declare PetalLayerLoop = 0;
#while (PetalLayerLoop < 4)
#declare PetalScale = 1-PetalLayerLoop / 5;
#declare PetalLoop = 0;
#while (PetalLoop < 3)
object
{
Petal
rotate <0, PetalLoop * 120 + PetalLayerLoop * 59, 0>
scale <PetalScale, 1, PetalScale>
}
#declare PetalLoop = PetalLoop + 1;
#end
#declare PetalLayerLoop = PetalLayerLoop + 1;
#end
texture
{
TulipHeadTexture
}
}
#declare TulipStem = intersection
{
torus
{
10, 0.06
rotate <90, 0, 0>
translate <10, 0, 0>
}
box
{
<-2, -5, -2>
<2, 0, 2>
}
texture
{
TulipStemTexture
}
}
#macro Tulip(HeadAngle)
union
{
object {TulipStem}
object {TulipHead rotate <0, HeadAngle, 0>}
//object{sphere {<0, 0, 0>, 0.5 texture{TulipHeadTexture}}}
}
#end
#declare Grass = difference
{
torus {
0.8, 0.1
}
torus {
0.8, 0.11
translate <0.02, 0, 0>
}
plane {
< 0, 0, -1>, 0
}
plane {
<-1, 0, 0>, 0
}
Cluster Computing
Page 115
Revision Version 2.3
Copyright © 2003
rotate <90, 0, 0>
translate <0.9, 0, 0>
scale <1, 2, 1>
texture
{
GrassTexture
}
bounded_by{box{<-0.08, 0, -0.1>, <0.13, 1.7, 0.1>} rotate <0, 0, -12>}
}
#declare WindmillBlades
{
box {< 2.0, -2.5, 0>,
box {< 2.5, 2.0, 0>,
box {<-2.0, 2.5, 0>,
box {<-2.5, -2.0, 0>,
= union
< 16.0,
0.5,
< -0.5, 16.0,
<-16.0, -0.5,
< 0.5, -16.0,
0.2>}
0.2>}
0.2>}
0.2>}
box {<-8.0, -0.5, 0>, <8.0, 0.5, 0.2>}
box {<-0.5, -8.0, 0>, <0.5, 8.0, 0.2>}
cylinder {<0, 0, 0>, <0, 0, 2>, 1}
texture
{
WindmillBladesTexture
}
}
#declare WindmillBase = union
{
cone
{
<0, 0, 0>, 8, <0, 20, 0>, 2
texture
{
WindmillBaseTexture
}
}
cylinder
cylinder
cylinder
cylinder
cylinder
{< 0,
{< 0,
{< 11,
{<-11,
{< 0,
-2,
0>,
-2,
0>,
-2,
0>,
-2,
0>,
-2, -11>,
< 0, 0.0,
0>, 12.0}
< 0, -8.0,
0>, 8.0}
< 11, 3.5,
0>, 0.2}
<-11, 3.5,
0>, 0.2}
< 0, 3.0, -11>, 0.2}
texture
{
WindmillBladesTexture
}
}
#declare Windmill = union
{
object
{
WindmillBase
}
object
{
WindmillBlades
rotate < 0, 0, -15>
rotate <15, 0,
0>
translate <0, 18, -6>
rotate <0, 20, 0>
}
}
#declare Sky = sky_sphere
{
Page 116
Cluster Computing
Revision Version 2.3
Copyright © 2003
pigment
{
bozo
turbulence 0.65
octaves 6
omega 0.7
lambda 2
color_map {
[0.0 color rgb<0.5, 0.5, 0.5>]
[1.0 color rgb<0.5, 0.7, 1.0>]
}
rotate <0, 30, 0>
rotate <73, 0, 0>
scale <0.2, 0.1, 0.1>
}
pigment
{
bozo
turbulence 0.65
octaves 6
omega 0.7
lambda 2
color_map {
[0.0 color rgbt<1.0, 1.00, 1, 0>]
[1.0 color rgbt<0.5, 0.66, 1, 0.2>]
}
rotate <0, 30, 0>
rotate <70, 0, 0>
scale <0.2, 0.1, 0.1>
}
}
camera
{
right x
up 4/3*y
location <0, 0, -3.0>
look_at <0, 0, 0>
aperture 0.25
blur_samples 256
focal_point<0,0,0>
confidence 0.9999
variance 0
}
#declare R1 = seed(123);
#declare LightLoop = 0;
#while (LightLoop < 100)
light_source
{
<800-rand(R1)*1600, 400, 800-rand(R1)*1600>
color rgb<0.037, 0.037, 0.037>
}
#declare LightLoop = LightLoop + 1;
#end
sky_sphere {Sky}
union
{
plane
{
<0, 1, 0>, -4
pigment {colour rgb <0, 0.5, 0>}
finish {ambient 1.0 diffuse 0.0}
}
Cluster Computing
Page 117
Revision Version 2.3
object {Tulip(65)
0.1, 0.0>}
object {Tulip(30)
1.2, 0.0>}
object {Tulip(70)
2.8, 3.8>}
object {Tulip(20)
1.2, 20.0>}
object {Tulip( 0)
2.0, 7.0>}
Copyright © 2003
rotate <0, 0, -25> rotate <0, -25, 0> translate <-0.7, rotate <0, 0, -40> rotate <0, -10, 0> translate <-0.2, rotate <0, 0, -15> rotate <0,
rotate <0, 0,
rotate <0, 0,
0, 0> translate < 1.8, -
0> rotate <0,
0> rotate <0,
0, 0> translate < 3.0,
0, 0> translate < 4.2, -
#declare R1 = seed(129);
#declare TulipLoop = 0;
#while (TulipLoop < 640)
object
{
Tulip(60+rand(R1)*360)
rotate <0, 0, -25-rand(R1)*10> // Lean
rotate <0, -rand(R1)*10, 0>
// Rotate
translate <rand(R1)*20, -rand(R1)*1, rand(R1)*40>
rotate <0, -45, 0>
rotate <-5, 0, 0>
translate <2, -0.7, 3>
}
#declare TulipLoop = TulipLoop + 1;
#end
#declare TulipLoop = 0;
#while (TulipLoop < 64)
object
{
Tulip(60+rand(R1)*360)
rotate <0, 0, -30-rand(R1)*10> // Lean
rotate <0, -rand(R1)*10, 0>
// Rotate
translate <-rand(R1)*20, -rand(R1)*1, rand(R1)*10>
rotate <0, -45, 0>
rotate <-5, 0, 0>
translate <-0.5, -1.5, 5>
}
#declare TulipLoop = TulipLoop + 1;
#end
#declare R1 = seed(696);
#declare GrassLoop = 0;
#while (GrassLoop < 4096)
object
{
Grass
rotate <-30 + 60*rand(R1), 0, -30 + 60*rand(R1)>
rotate <0, 360*rand(R1), 0>
scale <1 + rand(R1), 0.5 + rand(R1), 1 + rand(R1)>
translate <-10 + 20*rand(R1), -4, rand(R1)*20>
}
#declare GrassLoop = GrassLoop + 1;
#end
object
{
Windmill
translate <18, 9, 80>
}
#declare R1 = seed(60);
#declare GrassLoop = 0;
#while (GrassLoop < 48)
object
{
Grass
rotate <0, -140+100*rand(R1), 0>
scale <20, 1+1.5*rand(R1), 8>
translate <GrassLoop - 22, (48-GrassLoop)/24, 40>
}
Page 118
Cluster Computing
Revision Version 2.3
Copyright © 2003
#declare GrassLoop = GrassLoop + 1;
#end
#declare R1 = seed(19);
#declare GrassLoop = 0;
#while (GrassLoop < 128)
object
{
Grass
scale <2, (64 - abs(GrassLoop - 64))/64+1, 2>
rotate <0, -90, -90 + GrassLoop * 2>
translate <-20+GrassLoop/24+4*rand(R1), 2+6*rand(R1), 36>
}
#declare GrassLoop = GrassLoop + 1;
#end
}
8.4.3.
POV-Ray Benchmark Performance data
The following data is an extract of the POVVBENCH database detailing only those machines
that have significance to the testing done in the laboratory.
Note: All costs are in US dollars.
POVBench Result Extract
15-Jan-02
Found 1922 matches (serial & parallel)
Time
POVmark
Machine
Processor
0:00:01
14800
DARK STAR
866/24
Intel P-III
0:00:02
7400
MK-81
0:00:03
4933.33
0:00:04
3700
0:00:19
778.95
Monolith Cluster 4 X Dual
Celeron 500
0:00:19
778.95
Quad MMC
Daemon
0:00:20
740
0:00:21
0:00:21
Clock Rate
866 MHz
Operating
System
Compiler
POV
Version
Release
kernel
2.2.18_Cluster
gcc 2.95
Parallel: POV
3.1
May-01
Intel Pentium III 500 MHz
(96 total)
Linux 2.2.2
gcc
Parallel: POV
3.01
Jun-99
Fianna
dual intel pIII
450 MHz
redhat linux 6.2
egcs 2.91.66
Parallel: POV
3.1
May-01
FAST
25 x Intel
Celeron
525 MHz
Red Hat 6.1
egcs 2.91.66
Parallel: POV
3.1
Feb-00
500 MHz
FreeBSD 4.1.1- GCC
Stable
Parallel: POV 31
Jan-01
16 x Pentium
200 MHz
Pro 256k cache
Linux 2.1.127
Parallel: POV
3.02
Jan-99
Beowulf Test
Cluster
Intel Pentium II
(Cluster of 8
Machines)
233 MHz
Linux Redhat 7.2 GCC / MPICC
(2.4.7-10 Kernel
)
Parallel: POV
3.1
Jan-02
704.76
chloe
2x550 & 2x466
Celeron
550 MHz
SuSe Linux 6.3
Parallel: POV
3.1
Jan-00
704.76
SMILE Beowulf Pentium II
Cluster
350 MHz
Linux Redhat 5.2 GCC
Parallel: POV
3.02
Apr-99
Single: POV
3.02
Jun-98
gcc
egcs-2.91.66
Single Processor Results
0:02:08
115.63
Foremost
Pentium II
333 MHz
Microsoft NT
4.00.1381
??
0:02:09
114.73
Beowulf Test
Cluster Node 1
Only
Intel Pentium II
233 MHz
Linux Redhat 7.2 GCC
(2.4.2-10 Kernel)
Single: POV 3.1
Jan-02
0:02:11
112.98
bigone
PII-300
300 MHz
Linux 2.2.5 glibc egcs 1.1.2
Single: POV
3.02
May-99
0:02:11
112.98
Osborne
PowerATX PII
300
PII 300MHz /
64MB EDO
DRAM
300 MHz
Windows NT 4
(sp3)
Single: POVRay for
Windows 3.02
Nov-97
0:02:12
112.12
MaxCom PC
AMD-K6
350 MHz
Windows NT 4.0 ??
SP3
Single: POV 3.1
May-99
0:02:12
112.12
Compaq
Deskpro 6000
Pentium Pro
(256K)
200 MHz
Windows NT 4.0 Microsoft Visual Single: POV 2.2
SP3
C++ v5.0
Apr-97
??
Table 8-3 – POVBENCH POV-Ray Performance Benchmark Data Extract
Cluster Computing
Page 119
Revision Version 2.3
Copyright © 2003
8.5. Appendix E – Raw Results
8.5.1.
Lam MPI Cluster boot-up and tear-down
Script started on Tue Nov 20 22:47:31 2001
[beowulf@node1 ~]$ lamboot -v /etc/lam/lam-bhost.lam
LAM 6.5.5/MPI 2 C++/ROMIO - University of Notre Dame
Executing hboot on n0 (node1 - 1 CPU)...
Executing hboot on n1 (node2 - 1 CPU)...
Executing hboot on n2 (node3 - 1 CPU)...
Executing hboot on n3 (node4 - 1 CPU)...
Executing hboot on n4 (node5 - 1 CPU)...
Executing hboot on n5 (node6 - 1 CPU)...
Executing hboot on n6 (node7 - 1 CPU)...
Executing hboot on n7 (node8 - 1 CPU)...
topology n0(o)...n1...n2...n3...n4...n5...n6...n7...done
[beowulf@node1 ~]$ wipe -v /etc/lam/lam-bhost.lam
LAM 6.5.5/MPI 2 C++/ROMIO - University of Notre Dame
Executing tkill on
Executing tkill on
Executing tkill on
Executing tkill on
Executing tkill on
Executing tkill on
Executing tkill on
Executing tkill on
[beowulf@node1 ~]$
Script done on Tue
Page 120
n0 (node1)...
n1 (node2)...
n2 (node3)...
n3 (node4)...
n4 (node5)...
n5 (node6)...
n6 (node7)...
n7 (node8)...
exit
Nov 20 22:48:17 2001
Cluster Computing
Revision Version 2.3
8.5.2.
Copyright © 2003
POV-Ray
The following script files show the raw input/output from the command line. Each file has had
trivial data removed in part to reduce the overall size.
Skyvase on 1 node
Script started on Fri Nov 23 23:02:24 2001
[beowulf@node8 povray31]$ -__[Kx-povray -i skyvase.pov +v1 -d ___[K+ft -x +a0.300 +r3 -q9
w640 -h480 -mv2.0 +b1000 > results.txt Persistence of Vision(tm) Ray Tracer Version
3.1g.Linux.gcc
This is an official version prepared by the POV-Ray Team(tm). See the
documentation on how to contact the authors or visit us on the
internet at http://www.povray.org.
Copyright 1999 POV-Ray Team(tm)
Parsing Options
Input file: skyvase.pov (compatible to version 2.0)
Remove bounds........On Split unions........Off
Library paths: /usr/local/lib/povray31 /usr/local/lib/povray31/include
Output Options
Image resolution 640 by 480 (rows 1 to 480, columns 1 to 640).
Output file: skyvase.tga, 24 bpp Targa, 1000 KByte buffer
Graphic display.....Off
Mosaic preview......Off
CPU usage histogram.Off
Continued trace.....Off Allow interruption..Off Pause when done.....Off
Verbose messages.....On
Tracing Options
Quality: 9
Bounding boxes.......On Bounding threshold: 25
Light Buffer.........On Vista Buffer.........On
Antialiasing.........On (Method 1, Threshold 0.300, Depth 3, Jitter 1.00)
Radiosity...........Off
Animation Options
Clock value....
0.000 (Animation off)
Redirecting Options
All Streams to console..........On
Debug Stream to console.........On
Fatal Stream to console.........On
Render Stream to console........On
Statistics Stream to console....On
Warning Stream to console.......On
_-
Parsing...........skyvase.pov:83: warning: CSG union unnecessarily bounded.
Creating bounding slabs.
Scene contains 4 frame level objects; 3 infinite.
Rendering...
-:--:-- Rendering line
0 of 480Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
supersampled 0 times.
-:--:-- Rendering line
1 of 480 supersampled 0 times.
-:--:-- Rendering line
2 of 480 supersampled 0 times.
0:00:01 Rendering line
3 of 480 supersampled 0 times.
.
.
.
0:02:09 Rendering line 479 of 480 supersampled 10 times.
0:02:09 Rendering line 480 of 480 supersampled 10 times.
Done Tracing
skyvase.pov Statistics, Resolution 640 x 480
---------------------------------------------------------------------------Pixels:
307840
Samples:
396160
Smpls/Pxl: 1.29
Rays:
1519523
Saved:
9
Max Level: 5/5
---------------------------------------------------------------------------Ray->Shape Intersection
Tests
Succeeded Percentage
---------------------------------------------------------------------------Cone/Cylinder
2382542
1281629
53.79
CSG Intersection
3664171
396353
10.82
CSG Union
2563258
738033
28.79
Cluster Computing
Page 121
Revision Version 2.3
Copyright © 2003
Plane
26388678
14489601
54.91
Quadric
2563258
1571892
61.32
Sphere
2563258
652132
25.44
Bounding Object
2382542
1281629
53.79
---------------------------------------------------------------------------Calls to Noise:
1445139
Calls to DNoise:
2816194
---------------------------------------------------------------------------Shadow Ray Tests:
3726692
Succeeded:
90399
Reflected Rays:
1123363
---------------------------------------------------------------------------Smallest Alloc:
12 bytes
Largest:
1024008
Peak memory used:
1107606 bytes
---------------------------------------------------------------------------Time For Trace:
0 hours 2 minutes
9.0 seconds (129 seconds)
Total Time:
0 hours 2 minutes
9.0 seconds (129 seconds)
[beowulf@node8 povray31]$ x-povray -i skyvase.pov +v1 -d +ft -x +a0.300 +r3 -q9
-w640 -h480 -mv2.0 +b1000 > results.txt Persistence of Vision(tm) Ray Tracer Version
3.1g.Linux.gcc
This is an official version prepared by the POV-Ray Team(tm). See the
documentation on how to contact the authors or visit us on the
internet at http://www.povray.org.
Copyright 1999 POV-Ray Team(tm)
Parsing Options
Input file: skyvase.pov (compatible to version 2.0)
Remove bounds........On Split unions........Off
Library paths: /usr/local/lib/povray31 /usr/local/lib/povray31/include
Output Options
Image resolution 640 by 480 (rows 1 to 480, columns 1 to 640).
Output file: skyvase.tga, 24 bpp Targa, 1000 KByte buffer
Graphic display.....Off
Mosaic preview......Off
CPU usage histogram.Off
Continued trace.....Off Allow interruption..Off Pause when done.....Off
Verbose messages.....On
Tracing Options
Quality: 9
Bounding boxes.......On Bounding threshold: 25
Light Buffer.........On Vista Buffer.........On
Antialiasing.........On (Method 1, Threshold 0.300, Depth 3, Jitter 1.00)
Radiosity...........Off
Animation Options
Clock value....
0.000 (Animation off)
Redirecting Options
All Streams to console..........On
Debug Stream to console.........On
Fatal Stream to console.........On
Render Stream to console........On
Statistics Stream to console....On
Warning Stream to console.......On
Parsing...........skyvase.pov:83: warning: CSG union unnecessarily bounded.
Creating bounding slabs.
Scene contains 4 frame level objects; 3 infinite.
Rendering...
-:--:-- Rendering line
Fog and participating media
supersampled 0 times.
-:--:-- Rendering line
-:--:-- Rendering line
-:--:-- Rendering line
-:--:-- Rendering line
0:00:01 Rendering line
.
.
.
0 of 480Camera is inside a non-hollow object.
may not work as expected.
1
2
3
4
5
of
of
of
of
of
480
480
480
480
480
supersampled
supersampled
supersampled
supersampled
supersampled
0
0
0
0
0
times.
times.
times.
times.
times.
0:02:08 Rendering line 479 of 480 supersampled 10 times.
0:02:09 Rendering line 480 of 480 supersampled 10 times.
skyvase.pov Statistics, Resolution 640 x 480
Page 122
Done Tracing
Cluster Computing
Revision Version 2.3
Copyright © 2003
---------------------------------------------------------------------------Pixels:
307840
Samples:
396160
Smpls/Pxl: 1.29
Rays:
1519523
Saved:
9
Max Level: 5/5
---------------------------------------------------------------------------Ray->Shape Intersection
Tests
Succeeded Percentage
---------------------------------------------------------------------------Cone/Cylinder
2382542
1281629
53.79
CSG Intersection
3664171
396353
10.82
CSG Union
2563258
738033
28.79
Plane
26388678
14489601
54.91
Quadric
2563258
1571892
61.32
Sphere
2563258
652132
25.44
Bounding Object
2382542
1281629
53.79
---------------------------------------------------------------------------Calls to Noise:
1445139
Calls to DNoise:
2816194
---------------------------------------------------------------------------Shadow Ray Tests:
3726692
Succeeded:
90399
Reflected Rays:
1123363
---------------------------------------------------------------------------Smallest Alloc:
12 bytes
Largest:
1024008
Peak memory used:
1107606 bytes
---------------------------------------------------------------------------Time For Trace:
0 hours 2 minutes
9.0 seconds (129 seconds)
Total Time:
0 hours 2 minutes
9.0 seconds (129 seconds)
[beowulf@node8 povray31]$ exit
Script done on Fri Nov 23 23:08:30 2001
Cluster Computing
Page 123
Revision Version 2.3
Copyright © 2003
Skyvase on 2 nodes
Script started on Sat Nov 24 00:15:17 2001
[beowulf@node8 povray31]$ _mpirun n0 n0-1 ./mpi-x-povray -i skyvase.pov +v1 -d +f _t -x
+a0.300 +r3 -q9 -w640 -h480 -mv2.0 +b1000 > results.txt Persistence of Vision(tm) Ray Tracer
Version 3.1g
This is an unofficial version compiled by:
Leon Verrall ([email protected])
The POV-Ray Team(tm) is not responsible for supporting this version.
Copyright 1999 POV-Ray Team(tm)
Initializing MPI-POVRAY
Slave PE 2 successfully started.
STARTING FRAME 0...
Parsing (Slave PE 2)Slave PE 1 successfully started.
Parsing (Slave PE 1).............skyvase.pov:83: warning: CSG union unnecessarily bounded.
Creating bounding slabs.
Scene contains 4 frame level objects; 3 infinite.
Rendering...(Slave PE 2)
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
............skyvase.pov:83: warning: CSG union unnecessarily bounded.
.
..
Creating bounding slabs..
Scene contains 4 frame level objects; 3 infinite.
.
Rendering...(Slave PE 1)
.Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
........................
0.33 of blocks complete.
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
.............................
0.67 of blocks complete.
..........................Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
..
1.00 of blocks complete.
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
......................................
1.33 of blocks complete.
............Camera is inside a non-hollow object.
.
.
.
99.67 of blocks complete.
All blocks are assigned. Stopping PE 2.
Done Tracing
Slave PE 2 has exited. 1 minions left...
..........
99.75 of blocks complete... 456 of 480 lines finished (in frame 0)....
99.83 of blocks complete.. 464 of 480 lines finished (in frame 0)........
99.92 of blocks complete... 472 of 480 lines finished (in frame 0)......
100.00 of blocks complete.
100.00 of blocks complete. 480 of
480 lines finished (in frame 0).
Finishing Frame 0...
All blocks are assigned. Stopping PE 1.
Done Tracing
Page 124
Cluster Computing
Revision Version 2.3
Copyright © 2003
Slave PE 1 has exited. 0 minions left...
All slave tasks have exited!
PE Distribution Statistics:
Slave PE [ done ]
1 [42.00%]
2 [58.00%]
Slave PE
[ done ]
POV-Ray statistics for finished frames:
skyvase.pov Statistics (Partial Image Rendered), Resolution 640 x 480
---------------------------------------------------------------------------Pixels:
133056
Samples:
186768
Smpls/Pxl: 1.40
Rays:
738432
Saved:
0
Max Level: 0/5
---------------------------------------------------------------------------Ray->Shape Intersection
Tests
Succeeded Percentage
---------------------------------------------------------------------------Cone/Cylinder
1155707
633920
54.85
CSG Intersection
1789627
202500
11.32
CSG Union
1267840
383159
30.22
Plane
12824910
7000758
54.59
Quadric
1267840
796920
62.86
Sphere
1267840
336564
26.55
Bounding Object
1155707
633920
54.85
---------------------------------------------------------------------------Calls to Noise:
729155
Calls to DNoise:
1446800
---------------------------------------------------------------------------Shadow Ray Tests:
1812532
Succeeded:
45367
Reflected Rays:
551664
---------------------------------------------------------------------------Smallest Alloc:
32 bytes
Largest:
1024008
Peak memory used:
1720450 bytes
---------------------------------------------------------------------------Time For Trace:
0 hours 1 minutes 12.0 seconds (72 seconds)
Total Time:
0 hours 1 minutes 12.0 seconds (72 seconds)
[beowulf@node8 povray31]$ exit
Script done on Sat Nov 24 00:17:56 2001
Cluster Computing
Page 125
Revision Version 2.3
Copyright © 2003
Skyvase on 3 nodes
Script started on Sat Nov 24 00:22:09 2001
[beowulf@node8 povray31]$ mpirun no__[K0 n0-2 ./mpi-x-povray -i skyvase.pov +v1 -d +f _t -x
___[K+a0.300 +r3 -q9 -w640 -h480 -mv2.0 +b1000 > results.txt Persistence of Vision(tm) Ray
Tracer Version 3.1g
This is an unofficial version compiled by:
Leon Verrall ([email protected])
The POV-Ray Team(tm) is not responsible for supporting this version.
Copyright 1999 POV-Ray Team(tm)
Initializing MPI-POVRAY
Slave PE 2 successfully started.
Slave PE 3 successfully started.
STARTING FRAME 0...
Slave PE 1 successfully started.
Parsing (Slave PE 3).
Parsing (Slave PE 2)..
Parsing (Slave PE 1).................skyvase.pov:83: warning: CSG union unnecessarily bounded.
skyvase.pov:83: warning: CSG union unnecessarily bounded.
Creating bounding slabs.
Scene contains 4 frame level objects; 3 infinite.
Creating bounding slabs.
Scene contains 4 frame level objects; 3 infinite.
Rendering...(Slave PE 2)
Rendering...(Slave PE 3)
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
.....................................
0.33 of blocks complete.
....skyvase.pov:83: warning: CSG union unnecessarily bounded.
....
....
Creating bounding slabs...
Scene contains 4 frame level objects; 3 infinite.
..
Rendering...(Slave PE 1)
..Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
...........Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
..
0.67 of blocks complete.
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
......................................
1.00 of blocks complete.
......................................
1.33 of blocks complete.
.
.
.
99.33 of blocks complete.
.Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
....All blocks are assigned. Stopping PE 2.
Done Tracing
Slave PE 2 has exited. 2 minions left...
..........
99.50 of blocks complete. 456 of 480 lines finished (in frame 0)..................
99.67 of blocks complete.
.
99.67 of blocks complete.. 464 of 480 lines finished (in frame 0)..............All blocks are
assigned. Stopping PE 3.
Page 126
Cluster Computing
Revision Version 2.3
Copyright © 2003
Done Tracing
Slave PE 3 has exited. 1 minions left...
.......
99.92 of blocks complete.... 472 of 480 lines finished (in frame 0).
100.00 of blocks complete.
100.00 of blocks complete. 480 of
480 lines finished (in frame 0).
Finishing Frame 0...
All blocks are assigned. Stopping PE 1.
Done Tracing
Slave PE 1 has exited. 0 minions left...
All slave tasks have exited!
PE Distribution Statistics:
Slave PE [ done ]
1 [24.00%]
2 [37.33%]
Slave PE
[ done ]
3
[38.67%]
POV-Ray statistics for finished frames:
skyvase.pov Statistics (Partial Image Rendered), Resolution 640 x 480
---------------------------------------------------------------------------Pixels:
76032
Samples:
106488
Smpls/Pxl: 1.40
Rays:
432403
Saved:
0
Max Level: 0/5
---------------------------------------------------------------------------Ray->Shape Intersection
Tests
Succeeded Percentage
---------------------------------------------------------------------------Cone/Cylinder
670749
377745
56.32
CSG Intersection
1048494
123796
11.81
CSG Union
755490
227670
30.14
Plane
7462980
4090889
54.82
Quadric
755490
459300
60.79
Sphere
755490
199338
26.39
Bounding Object
670749
377745
56.32
---------------------------------------------------------------------------Calls to Noise:
459305
Calls to DNoise:
889040
---------------------------------------------------------------------------Shadow Ray Tests:
1048064
Succeeded:
28543
Reflected Rays:
325915
---------------------------------------------------------------------------Smallest Alloc:
32 bytes
Largest:
1024008
Peak memory used:
1822934 bytes
---------------------------------------------------------------------------Time For Trace:
0 hours 0 minutes 48.0 seconds (48 seconds)
Total Time:
0 hours 0 minutes 48.0 seconds (48 seconds)
[beowulf@node8 povray31]$ exit
Script done on Sat Nov 24 00:24:35 2001
Cluster Computing
Page 127
Revision Version 2.3
Copyright © 2003
Skyvase on 4 nodes
Script started on Sun Nov 25 18:11:01 2001
[beowulf@node8 povray31]$ mpirun n0 n0-3 ./p__[Kmpi-x-povray -i sj__[Kkyvase.pov +v1 -d +f _t
-x +a0.300 +r3 -q9 -w640 -h48-__[K0 -mv2.0 +b1000 > results.txt Persistence of Vision(tm) Ray
Tracer Version 3.1g
This is an unofficial version compiled by:
Leon Verrall ([email protected])
The POV-Ray Team(tm) is not responsible for supporting this version.
Copyright 1999 POV-Ray Team(tm)
Initializing MPI-POVRAY
Slave PE 2 successfully started.
Slave PE 4 successfully started.
STARTING FRAME 0...
Slave PE 1 successfully started.
Slave PE 3 successfully started.
Parsing (Slave PE 4).
Parsing (Slave PE 1)....
Parsing (Slave PE 2).....
Parsing (Slave PE 3).................skyvase.pov:83: warning: CSG union unnecessarily bounded.
.skyvase.pov:83: warning: CSG union unnecessarily bounded.
Creating bounding slabs.
Scene contains 4 frame level objects; 3 infinite.
skyvase.pov:83: warning: CSG union unnecessarily bounded.
Creating bounding slabs.
Scene contains 4 frame level objects; 3 infinite.
Creating bounding slabs.
Rendering...(Slave PE 2)
Rendering...(Slave PE 4)
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
Scene contains 4 frame level objects; 3 infinite.
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
Rendering...(Slave PE 3)
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
..............................................
0.33 of blocks complete.
.............................
0.67 of blocks complete.
.............skyvase.pov:83: warning: CSG union unnecessarily bounded.
......Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
....Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
..
1.00 of blocks complete.
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
.
....
Creating bounding slabs.......
Scene contains 4 frame level objects; 3 infinite.
......
Rendering...(Slave PE 1)
......Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
.....................
1.33 of blocks complete.
Page 128
Cluster Computing
Revision Version 2.3
Copyright © 2003
99.33 of blocks complete.
......Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
........All blocks are assigned. Stopping PE 2.
Done Tracing..
Slave PE 2 has exited. 3 minions left...
.........All blocks are assigned. Stopping PE 4.
.
Done Tracing
Slave PE 4 has exited. 2 minions left...
.
99.67 of blocks complete.
..
99.67 of blocks complete. 456 of 480 lines finished (in frame 0)...........
99.83 of blocks complete. 464 of 480 lines finished (in frame 0).All blocks are assigned.
Stopping PE 1.
.
Done Tracing.
Slave PE 1 has exited. 1 minions left...
......
99.92 of blocks complete. 472 of 480 lines finished (in frame 0).......
100.00 of blocks complete.
100.00 of blocks complete. 480 of
480 lines finished (in frame 0).
Finishing Frame 0...
All blocks are assigned. Stopping PE 3.
Done Tracing
Slave PE 3 has exited. 0 minions left...
All slave tasks have exited!
PE Distribution Statistics:
Slave PE [ done ]
1 [15.00%]
2 [29.33%]
4 [28.00%]
Slave PE
[ done ]
3
[27.67%]
POV-Ray statistics for finished frames:
skyvase.pov Statistics (Partial Image Rendered), Resolution 640 x 480
---------------------------------------------------------------------------Pixels:
87648
Samples:
116560
Smpls/Pxl: 1.33
Rays:
454282
Saved:
3
Max Level: 0/5
---------------------------------------------------------------------------Ray->Shape Intersection
Tests
Succeeded Percentage
---------------------------------------------------------------------------Cone/Cylinder
713545
379049
53.12
CSG Intersection
1092594
119175
10.91
CSG Union
758098
217557
28.70
Plane
7893548
4307876
54.57
Quadric
758098
480966
63.44
Sphere
758098
190100
25.08
Bounding Object
713545
379049
53.12
---------------------------------------------------------------------------Calls to Noise:
430048
Calls to DNoise:
840608
---------------------------------------------------------------------------Shadow Ray Tests:
1118580
Succeeded:
24834
Reflected Rays:
337722
---------------------------------------------------------------------------Smallest Alloc:
32 bytes
Largest:
1024008
Peak memory used:
1925418 bytes
---------------------------------------------------------------------------Time For Trace:
0 hours 0 minutes 36.0 seconds (36 seconds)
Total Time:
0 hours 0 minutes 36.0 seconds (36 seconds)
[beowulf@node8 povray31]$ exit
Script done on Sun Nov 25 18:13:10 2001
Cluster Computing
Page 129
Revision Version 2.3
Copyright © 2003
Skyvase on 5 nodes
Script started on Sat Nov 24 00:29:40 2001
[beowulf@node8 povray31]$ mpirun n0 n0-4 mpirun__[K__[K__[K__[K__[Kpi-x-povray -i skyvase.pov
+v1 -d +ft _-x +a0.300 +r3 -q9 -w640 -h480 -mv2.0 +b1000 > results.txt Persistence of
Vision(tm) Ray Tracer Version 3.1g
This is an unofficial version compiled by:
Leon Verrall ([email protected])
The POV-Ray Team(tm) is not responsible for supporting this version.
Copyright 1999 POV-Ray Team(tm)
Initializing MPI-POVRAY
Slave PE 2 successfully started.
Slave PE 4 successfully started.
STARTING FRAME 0...
Parsing (Slave PE 4)Slave PE 3 successfully started.
Slave PE 5 successfully started.
.
Parsing (Slave PE 5).
Parsing (Slave PE 2)....
Parsing (Slave PE 3)...Slave PE 1 successfully started.
.
Parsing (Slave PE 1)....................skyvase.pov:83: warning: CSG union unnecessarily
bounded.
.skyvase.pov:83: warning: CSG union unnecessarily bounded.
Creating bounding slabs.
.
Scene contains 4 frame level objects; 3 infinite.
.
Rendering...(Slave PE 2)
Creating bounding slabs.
Scene contains 4 frame level objects; 3 infinite.
..skyvase.pov:83: warning: CSG union unnecessarily bounded.
skyvase.pov:83: warning: CSG union unnecessarily bounded.
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
Rendering...(Slave PE 4)
.
.Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
.
Creating bounding slabs.
Creating bounding slabs.
Scene contains 4 frame level objects; 3 infinite.
Scene contains 4 frame level objects; 3 infinite.
.
Rendering...(Slave PE 3)
Rendering...(Slave PE 5)
.Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
....................................
0.33 of blocks complete.
................................
Rendering...(Slave PE 1)
98.67 of blocks complete.
.Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
............Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
...........All blocks are assigned. Stopping PE 1.
....
Done Tracing..
99.00 of blocks complete.
Page 130
Cluster Computing
Revision Version 2.3
Copyright © 2003
..
Slave PE 1 has exited. 4 minions left...
................................
99.25 of blocks complete. 456 of 480 lines finished (in frame 0)..
99.33 of blocks complete.
............................
99.50 of blocks complete. 464 of 480 lines finished (in frame 0)...
99.67 of blocks complete.
All blocks are assigned. Stopping PE 4.
..All blocks are assigned. Stopping PE 5.
..
Done Tracing..
Done Tracing..
Slave PE 4 has exited. 3 minions left...
......
99.83 of blocks complete. 472 of 480 lines finished (in frame 0).
Slave PE 5 has exited. 2 minions left...
.....All blocks are assigned. Stopping PE 2.
.
Done Tracing
Slave PE 2 has exited. 1 minions left...
...
100.00 of blocks complete.
100.00 of blocks complete. 480 of 480 lines finished (in frame 0).
Finishing Frame 0...
All blocks are assigned. Stopping PE 3.
Done Tracing
Slave PE 3 has exited. 0 minions left...
All slave tasks have exited!
PE Distribution Statistics:
Slave PE [ done ]
1 [11.00%]
2 [23.67%]
4 [21.00%]
Slave PE
[ done ]
3
5
[22.67%]
[21.67%]
POV-Ray statistics for finished frames:
skyvase.pov Statistics (Partial Image Rendered), Resolution 640 x 480
---------------------------------------------------------------------------Pixels:
71808
Samples:
95928
Smpls/Pxl: 1.34
Rays:
366381
Saved:
0
Max Level: 0/5
---------------------------------------------------------------------------Ray->Shape Intersection
Tests
Succeeded Percentage
---------------------------------------------------------------------------Cone/Cylinder
571431
297967
52.14
CSG Intersection
869398
99074
11.40
CSG Union
595934
176956
29.69
Plane
6310244
3463621
54.89
Quadric
595934
370114
62.11
Sphere
595934
155124
26.03
Bounding Object
571431
297967
52.14
---------------------------------------------------------------------------Calls to Noise:
355777
Calls to DNoise:
678932
---------------------------------------------------------------------------Shadow Ray Tests:
894800
Succeeded:
24286
Reflected Rays:
270453
---------------------------------------------------------------------------Smallest Alloc:
32 bytes
Largest:
1024008
Peak memory used:
1822974 bytes
---------------------------------------------------------------------------Time For Trace:
0 hours 0 minutes 29.0 seconds (29 seconds)
Total Time:
0 hours 0 minutes 29.0 seconds (29 seconds)
[beowulf@node8 povray31]$ exit
Script done on Sat Nov 24 00:31:43 2001
Cluster Computing
Page 131
Revision Version 2.3
Copyright © 2003
Skyvase on 6 nodes
Script started on Sat Nov 24 00:32:31 2001
[beowulf@node8 povray31]$ mpirun n0 n0-5 ./mpi-x-povray -i skyvase.pov +v1 -d +f _t -x +a0.300
+r3 -q9 -w640 -h480 -mv2.0 +b1000 > results.txt Persistence of Vision(tm) Ray Tracer Version
3.1g
This is an unofficial version compiled by:
Leon Verrall ([email protected])
The POV-Ray Team(tm) is not responsible for supporting this version.
Copyright 1999 POV-Ray Team(tm)
Initializing MPI-POVRAY
Slave PE 2 successfully started.
Slave PE 4 successfully started.
Slave PE 6 successfully started.
STARTING FRAME 0...
Slave PE 1 successfully started.
Slave PE 3 successfully started.
Slave PE 5 successfully started.
Parsing (Slave PE 1).
Parsing (Slave PE 6)
Parsing (Slave PE 2)
Parsing (Slave PE 3)
Parsing (Slave PE 5)....
Parsing (Slave PE 4)....................................skyvase.pov:83: warning: CSG union
unnecessarily bounded.
.
skyvase.pov:83: warning: CSG union unnecessarily bounded.
Creating bounding slabs.
Scene contains 4 frame level objects; 3 infinite.
Rendering...(Slave PE 2)
Creating bounding slabs.
Scene contains 4 frame level objects; 3 infinite.
..Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
skyvase.pov:83: warning: CSG union unnecessarily bounded.
skyvase.pov:83: warning: CSG union unnecessarily bounded.
.
Rendering...(Slave PE 4)
.Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
.
Creating bounding slabs.
Creating bounding slabs..
Scene contains 4 frame level objects; 3 infinite.
Scene contains 4 frame level objects; 3 infinite.
Rendering...(Slave PE 3)
Rendering...(Slave PE 5)
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
skyvase.pov:83: warning: CSG union unnecessarily bounded.
.Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
..
Creating bounding slabs..
Scene contains 4 frame level objects; 3 infinite.
Rendering...(Slave PE 6)
.......Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
....................
Page 132
Cluster Computing
Revision Version 2.3
Copyright © 2003
0.33 of blocks complete.
........................................
0.67 of blocks complete.
.....................................
1.00 of blocks complete.
.....................
1.33 of blocks complete.
...
Rendering...(Slave PE 1)
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
...Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
..........Camera is inside a non-hollow object.
99.67 of blocks complete.
..All blocks are assigned. Stopping PE 2.
.All blocks are assigned. Stopping PE 5.
Done TracingAll blocks are assigned. Stopping PE 4.
Slave PE 2 has exited. 3 minions left...
Done Tracing
Done Tracing
Slave PE 5 has exited. 2 minions left...
Slave PE 4 has exited. 1 minions left...
........
99.83 of blocks complete... 464 of 480 lines finished (in frame 0).....
99.92 of blocks complete.. 472 of 480 lines finished (in frame 0).......
100.00 of blocks complete.
100.00 of blocks complete. 480 of
480 lines finished (in frame 0).
Finishing Frame 0...
All blocks are assigned. Stopping PE 1.
Done Tracing
Slave PE 1 has exited. 0 minions left...
All slave tasks have exited!
PE Distribution Statistics:
Slave PE [ done ]
1 [ 8.67%]
2 [18.00%]
4 [19.00%]
6 [17.67%]
Slave PE
[ done ]
3
5
[18.33%]
[18.33%]
POV-Ray statistics for finished frames:
skyvase.pov Statistics (Partial Image Rendered), Resolution 640 x 480
---------------------------------------------------------------------------Pixels:
27456
Samples:
40240
Smpls/Pxl: 1.47
Rays:
155069
Saved:
3
Max Level: 0/5
---------------------------------------------------------------------------Ray->Shape Intersection
Tests
Succeeded Percentage
---------------------------------------------------------------------------Cone/Cylinder
245864
127244
51.75
CSG Intersection
373108
36704
9.84
CSG Union
254488
73191
28.76
Plane
2713128
1480394
54.56
Quadric
254488
162635
63.91
Sphere
254488
75631
29.72
Bounding Object
245864
127244
51.75
---------------------------------------------------------------------------Calls to Noise:
150696
Calls to DNoise:
291116
---------------------------------------------------------------------------Shadow Ray Tests:
387520
Succeeded:
6545
Cluster Computing
Page 133
Revision Version 2.3
Copyright © 2003
Reflected Rays:
114829
---------------------------------------------------------------------------Smallest Alloc:
32 bytes
Largest:
1024008
Peak memory used:
1925458 bytes
---------------------------------------------------------------------------Time For Trace:
0 hours 0 minutes 25.0 seconds (25 seconds)
Total Time:
0 hours 0 minutes 25.0 seconds (25 seconds)
[beowulf@node8 povray31]$ exit
Script done on Sat Nov 24 00:34:26 2001
Skyvase on 7 nodes
Script started on Fri Nov 23 23:13:37 2001
[beowulf@node8 povray31]$ _mpirun n0 N ./mpi-x-povray -i skyvase.pov +v1 -d +ft - _x +a0.300
+r3 -q9 -w640 -h480 -mv2.0 ___[K+b1000 > results.txt Persistence of Vision(tm) Ray Tracer
Version 3.1g
This is an unofficial version compiled by:
Leon Verrall ([email protected])
The POV-Ray Team(tm) is not responsible for supporting this version.
Copyright 1999 POV-Ray Team(tm)
Initializing MPI-POVRAY
Slave PE 2 successfully started.
Slave PE 4 successfully started.
Slave PE 6 successfully started.
STARTING FRAME 0...
Parsing (Slave PE 2)
Parsing (Slave PE 4)..Slave PE 1 successfully started.
Slave PE 3 successfully started.
Slave PE 5 successfully started.
Slave PE 7 successfully started.
..
Parsing (Slave PE 7)..
Parsing (Slave PE 1)....
Parsing (Slave PE 6).........
Parsing (Slave PE 3)..........
Parsing (Slave PE 5)............skyvase.pov:83: warning: CSG union unnecessarily bounded.
..
skyvase.pov:83: warning: CSG union unnecessarily bounded.
.
Creating bounding slabs.
Scene contains 4 frame level objects; 3 infinite.
Rendering...(Slave PE 2)
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
Creating bounding slabs.
Scene contains 4 frame level objects; 3 infinite.
Creating bounding slabs.
Rendering...(Slave PE 6)
Scene contains 4 frame level objects; 3 infinite.
Rendering...(Slave PE 4)
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
..........skyvase.pov:83: warning: CSG union unnecessarily bounded.
skyvase.pov:83: warning: CSG union unnecessarily bounded.
.
skyvase.pov:83: warning: CSG union unnecessarily bounded.
.
Creating bounding slabs..
Creating bounding slabs.
Scene contains 4 frame level objects; 3 infinite.
Scene contains 4 frame level objects; 3 infinite.
Page 134
Cluster Computing
Revision Version 2.3
Copyright © 2003
Rendering...(Slave PE 3)
Rendering...(Slave PE 5)
.
Creating bounding slabs.
Scene contains 4 frame level objects; 3 infinite.
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
..
Rendering...(Slave PE 7)
Camera is inside a non-hollow object.
Fog and participating media may not work as expected.
...................................................
0.33 of blocks complete.
..................
0.67 of blocks complete.
...................................
1.00 of blocks complete.
..........................................
1.33 of blocks complete.
98.67 of blocks complete.
...............All blocks are assigned. Stopping PE 6.
Done Tracing
Slave PE 6 has exited. 6 minions left...
........................
99.00 of blocks complete.
.....................
99.17 of blocks complete. 456 of 480 lines finished (in frame 0).......
99.33 of blocks complete.
....All blocks are assigned. Stopping PE 5.
....
Done Tracing...All blocks are assigned. Stopping PE 7.
...........
99.50 of blocks complete. 464 of 480 lines finished (in frame 0).
Slave PE 5 has exited. 5 minions left...
...
99.67 of blocks complete.
Done TracingAll blocks are assigned. Stopping PE 2.
.All blocks are assigned. Stopping PE 4.
Slave PE 7 has exited. 4 minions left...
Done Tracing
Done Tracing
Slave PE 2 has exited. 3 minions left...
Slave PE 4 has exited. 2 minions left...
.........All blocks are assigned. Stopping PE 3.
Done Tracing
Slave PE 3 has exited. 1 minions left...
...........
99.92 of blocks complete... 472 of 480 lines finished (in frame 0).
100.00 of blocks complete.
100.00 of blocks complete. 480 of
480 lines finished (in frame 0).
Finishing Frame 0...
All blocks are assigned. Stopping PE 1.
Done Tracing
Slave PE 1 has exited. 0 minions left...
All slave tasks have exited!
Cluster Computing
Page 135
Revision Version 2.3
PE Distribution Statistics:
Slave PE [ done ]
1 [ 6.33%]
2 [15.33%]
4 [15.33%]
6 [15.00%]
Copyright © 2003
Slave PE
[ done ]
3
5
7
[16.00%]
[15.67%]
[16.33%]
POV-Ray statistics for finished frames:
skyvase.pov Statistics (Partial Image Rendered), Resolution 640 x 480
---------------------------------------------------------------------------Pixels:
20064
Samples:
28008
Smpls/Pxl: 1.40
Rays:
104286
Saved:
0
Max Level: 0/5
---------------------------------------------------------------------------Ray->Shape Intersection
Tests
Succeeded Percentage
---------------------------------------------------------------------------Cone/Cylinder
161648
97365
60.23
CSG Intersection
259013
35699
13.78
CSG Union
194730
64334
33.04
Plane
1811210
993641
54.86
Quadric
194730
158522
81.41
Sphere
194730
65474
33.62
Bounding Object
161648
97365
60.23
---------------------------------------------------------------------------Calls to Noise:
122814
Calls to DNoise:
240434
---------------------------------------------------------------------------Shadow Ray Tests:
246972
Succeeded:
7013
Reflected Rays:
76278
---------------------------------------------------------------------------Smallest Alloc:
32 bytes
Largest:
1024008
Peak memory used:
1925478 bytes
---------------------------------------------------------------------------Time For Trace:
0 hours 0 minutes 22.0 seconds (22 seconds)
Total Time:
0 hours 0 minutes 22.0 seconds (22 seconds)
[beowulf@node8 povray31]$ exit
Script done on Fri Nov 23 23:15:47 2001
Skyvase on 8 nodes
Script started on Sun Nov 25 18:14:22 2001
[beowulf@node8 povray31]$ mpirun n0 n0-7 ./pm__[Kmpi__[K__[K__[K__[Kmpi-x-povray -i
skyvase.pov +v1 -d +f _t -x +a0.300 +r3 -q9 -w640 =h480__[K__[K__[K__[K__[K-h480 -mv2.0 +b1000
> results.txt Persistence of Vision(tm) Ray Tracer Version 3.1g
This is an unofficial version compiled by:
Leon Verrall ([email protected])
The POV-Ray Team(tm) is not responsible for supporting this version.
Copyright 1999 POV-Ray Team(tm)
Initializing MPI-POVRAY
Slave PE 1 successfully started.
Slave PE 2 successfully started.
Slave PE 3 successfully started.
Slave PE 4 successfully started.
Slave PE 5 successfully started.
Slave PE 6 successfully started.
Slave PE 7 successfully started.
Slave PE 8 successfully started.
STARTING FRAME 0...
Parsing (Slave PE 4)
Parsing (Slave PE 6)
Parsing (Slave PE 8)
Parsing (Slave PE 1)
Parsing (Slave PE 2)
Parsing (Slave PE 5)........
Parsing (Slave PE 3)..........
Parsing (Slave PE 7).......................................
skyvase.pov:83: warning: CSG union unnecessarily bounded.
Creating bounding slabs.
Scene contains 4 frame level objects; 3 infinite.
Page 136
Cluster Computing
Revision Version 2.3
Copyright © 2003
Rendering...(Slave PE 3)
Rendering...(Slave PE 2)
Rendering...(Slave PE 4)
Rendering...(Slave PE 8)
Rendering...(Slave PE 6)
Rendering...(Slave PE 7)
Rendering...(Slave PE 5)
..........................................................
0.33 of blocks complete.
.......................
0.67 of blocks complete.
1.00 of blocks complete.
..........................
1.33 of blocks complete.
Rendering...(Slave PE 1)
..........................
98.58 of blocks complete........... 456 of
98.67 of blocks complete.
...All blocks are assigned. Stopping PE 7.
480 lines finished (in frame 0)....
Done Tracing
Slave PE 7 has exited. 7 minions left...
.......................All blocks are assigned. Stopping PE 5.
Done Tracing..
Slave PE 5 has exited. 6 minions left...
...
99.00 of blocks complete.
..................................
99.33 of blocks complete.
....All blocks are assigned. Stopping PE 8.
...All blocks are assigned. Stopping PE 2.
...
Done Tracing..
Done Tracing...All blocks are assigned. Stopping PE 3.
.
Slave PE 8 has exited. 5 minions left...
.
Slave PE 2 has exited. 4 minions left...
Done Tracing
99.67 of blocks complete.
Slave PE 3 has exited. 3 minions left...
All blocks are assigned. Stopping PE 4.
Done Tracing
Slave PE 4 has exited. 2 minions left...
..All blocks are assigned. Stopping PE 6.
Done Tracing
Slave PE 6 has exited. 1 minions left...
....
99.83 of blocks complete. 464 of 480 lines finished (in frame 0)........
99.92 of blocks complete. 472 of 480 lines finished (in frame 0)........
100.00 of blocks complete.
100.00 of blocks complete. 480 of
480 lines finished (in frame 0).
Finishing Frame 0...
All blocks are assigned. Stopping PE 1.
Done Tracing
Slave PE 1 has exited. 0 minions left...
All slave tasks have exited!
PE Distribution Statistics:
Slave PE [ done ]
1 [ 4.67%]
Cluster Computing
Slave PE
[ done ]
Page 137
Revision Version 2.3
Copyright © 2003
2
4
6
8
[13.33%]
[13.67%]
[13.33%]
[14.00%]
3
5
7
[13.67%]
[14.33%]
[13.00%]
POV-Ray statistics for finished frames:
skyvase.pov Statistics (Partial Image Rendered), Resolution 640 x 480
---------------------------------------------------------------------------Pixels:
14784
Samples:
23888
Smpls/Pxl: 1.62
Rays:
81064
Saved:
0
Max Level: 0/5
---------------------------------------------------------------------------Ray->Shape Intersection
Tests
Succeeded Percentage
---------------------------------------------------------------------------Cone/Cylinder
123967
82687
66.70
CSG Intersection
206654
27632
13.37
CSG Union
165374
47952
29.00
Plane
1405044
779425
55.47
Quadric
165374
128278
77.57
Sphere
165374
43523
26.32
Bounding Object
123967
82687
66.70
---------------------------------------------------------------------------Calls to Noise:
91622
Calls to DNoise:
186122
---------------------------------------------------------------------------Shadow Ray Tests:
186268
Succeeded:
5234
Reflected Rays:
57176
---------------------------------------------------------------------------Smallest Alloc:
32 bytes
Largest:
1024008
Peak memory used:
2130426 bytes
---------------------------------------------------------------------------Time For Trace:
0 hours 0 minutes 20.0 seconds (20 seconds)
Total Time:
0 hours 0 minutes 20.0 seconds (20 seconds)
[beowulf@node8 povray31]$ exit
Script done on Sun Nov 25 18:16:23 2001
Poolballs on 1 node – Sequential POV-Ray
Script started on Sat Nov 24 03:03:00 2001
[beowulf@node8 povray31]$ x-povray -i demo.pov -w1024 -768___h768___ Persistence of Vision(tm)
Ray Tracer Version 3.1g.Linux.gcc
This is an official version prepared by the POV-Ray Team(tm). See the
documentation on how to contact the authors or visit us on the
internet at http://www.povray.org.
Copyright 1999 POV-Ray Team(tm)
Parsing Options
Input file: demo.pov (compatible to version 3.1)
Remove bounds........On Split unions........Off
Library paths: /usr/local/lib/povray31 /usr/local/lib/povray31/include
Output Options
Image resolution 1024 by 768 (rows 1 to 768, columns 1 to 1024).
Output file: demo.png, 24 bpp PNG
Graphic display.....Off
Mosaic preview......Off
CPU usage histogram.Off
Continued trace.....Off Allow interruption...On Pause when done.....Off
Verbose messages....Off
Tracing Options
Quality: 9
Bounding boxes.......On Bounding threshold: 25
Light Buffer.........On Vista Buffer.........On
Antialiasing........Off
Radiosity...........Off
Animation Options
Clock value....
0.000 (Animation off)
Redirecting Options
All Streams to console..........On
Debug Stream to console.........On
Fatal Stream to console.........On
Render Stream to console........On
Statistics Stream to console....On
Warning Stream to console.......On
Page 138
Cluster Computing
Revision Version 2.3
Copyright © 2003
Parsing........demo.pov:26: warning: All #version and #declares of float, vector, and color
require semi-colon ';' at end.
demo.pov:50: warning: All #version and #declares of float, vector, and color require semicolon ';' at end.
13_Pball.inc:100: warning: All #version and #declares of float, vector, and color require
semi-colon ';' at end.
No pigment type given.
Creating bounding slabs.
Scene contains 13 frame level objects; 0 infinite.
Rendering... Done Tracing
demo.pov Statistics, Resolution 1024 x 768
---------------------------------------------------------------------------Pixels:
786432
Samples:
786432
Smpls/Pxl: 1.00
Rays:
1343159
Saved:
3401
Max Level: 4/5
---------------------------------------------------------------------------Ray->Shape Intersection
Tests
Succeeded Percentage
---------------------------------------------------------------------------Box
19451664
13021308
66.94
Cone/Cylinder
25935552
5203596
20.06
CSG Intersection
95097024
4222255
4.44
CSG Union
34580736
1595749
4.61
Sphere
43225920
6257460
14.48
True Type Font
51871104
1392782
2.69
---------------------------------------------------------------------------Calls to Noise:
0
Calls to DNoise:
356028
---------------------------------------------------------------------------Shadow Ray Tests:
10679994
Succeeded:
92606
Reflected Rays:
556727
---------------------------------------------------------------------------Smallest Alloc:
10 bytes
Largest:
20508
Peak memory used:
179004 bytes
---------------------------------------------------------------------------Time For Trace:
0 hours 42 minutes 21.0 seconds (2541 seconds)
Total Time:
0 hours 42 minutes 21.0 seconds (2541 seconds)
[beowulf@node8 povray31]$ exit
Script done on Sat Nov 24 03:47:30 2001
Poolballs on 1 node – Parallel POV-Ray
Script started on Sun Nov 25 18:32:46 2001
[beowulf@node8 povray31]$ mpirun ./povray_________ ./po___[1@mp_p_[1@po___[1@ip__[1@0p__[1@p___[Pp___[Pp__[1@-p__[1@xp__[1@-p__________[1@n.__[1@0.__[1@ .__[1@n.__[1@0.__[1@ ._./mpi-xpovra y__y_[K_y -i demo.pov -h__[Kw1024 -h768 Persistence of Vision(tm) Ray Tracer Version
3.1g
This is an unofficial version compiled by:
Leon Verrall ([email protected])
The POV-Ray Team(tm) is not responsible for supporting this version.
Copyright 1999 POV-Ray Team(tm)
Initializing MPI-POVRAY
Parsing (Slave PE 1).....demo.pov:26: warning: All #version and #declares of float, vector,
and color require semi-colon ';' at end.
demo.pov:28: warning: All #version and #declares of float, vector, and color require semicolon ';' at end.
13_Pball.inc:98: warning: All #version and #declares of float, vector, and color require semicolon ';' at end.
No pigment type given.
Creating bounding slabs.
Scene contains 13 frame level objects; 0 infinite.
Rendering...(Slave PE 1)
...All blocks are assigned. Stopping PE 1.
Done Tracing
Slave PE 1 has exited. 0 minions left...
All slave tasks have exited!
PE Distribution Statistics:
Slave PE [ done ]
Cluster Computing
Slave PE
[ done ]
Page 139
Revision Version 2.3
Copyright © 2003
1
[100.00%]
POV-Ray statistics for finished frames:
demo.pov Statistics, Resolution 1024 x 768
---------------------------------------------------------------------------Pixels:
786432
Samples:
786432
Smpls/Pxl: 1.00
Rays:
1343159
Saved:
3401
Max Level: 0/5
---------------------------------------------------------------------------Ray->Shape Intersection
Tests
Succeeded Percentage
---------------------------------------------------------------------------Box
19451664
13021308
66.94
Cone/Cylinder
25935552
5203596
20.06
CSG Intersection
95097024
4222255
4.44
CSG Union
34580736
1595749
4.61
Sphere
43225920
6256890
14.47
True Type Font
51871104
1392782
2.69
---------------------------------------------------------------------------Calls to Noise:
0
Calls to DNoise:
356038
---------------------------------------------------------------------------Shadow Ray Tests:
10679994
Succeeded:
92606
Reflected Rays:
556727
---------------------------------------------------------------------------Smallest Alloc:
32 bytes
Largest:
20508
Peak memory used:
761161 bytes
---------------------------------------------------------------------------Time For Trace:
0 hours 29 minutes 46.0 seconds (1786 seconds)
Total Time:
0 hours 29 minutes 46.0 seconds (1786 seconds)
[beowulf@node8 povray31]$ exit
Script done on Sun Nov 25 19:04:38 2001
Poolballs on 2 nodes
Script started on Sat Nov 24 02:34:04 2001
[beowulf@node8 povray31]$ mpirun n0 n0-1 ./mpi-x-povray -i demo.pov -w6__[K1024 -h768 _
Persistence of Vision(tm) Ray Tracer Version 3.1g
This is an unofficial version compiled by:
Leon Verrall ([email protected])
The POV-Ray Team(tm) is not responsible for supporting this version.
Copyright 1999 POV-Ray Team(tm)
Initializing MPI-POVRAY
Parsing (Slave PE 1)...
Parsing (Slave PE 2).....demo.pov:26: warning: All #version and #declares of float, vector,
and color require semi-colon ';' at end.
demo.pov:28: warning: All #version and #declares of float, vector, and color require semicolon ';' at end.
13_Pball.inc:101: warning: All #version and #declares of float, vector, and color require
semi-colon ';' at end.
No pigment type given.
Creating bounding slabs.
Scene contains 13 frame level objects; 0 infinite.
Rendering...(Slave PE 2)
Creating bounding slabs.
Scene contains 13 frame level objects; 0 infinite.
Rendering...(Slave PE 1)
...All blocks are assigned. Stopping PE 2.
Done Tracing
Slave PE 2 has exited. 1 minions left...
......................All blocks are assigned. Stopping PE 1.
Done Tracing
Slave PE 1 has exited. 0 minions left...
All slave tasks have exited!
Page 140
Cluster Computing
Revision Version 2.3
PE Distribution Statistics:
Slave PE [ done ]
1 [48.96%]
2 [51.04%]
Copyright © 2003
Slave PE
[ done ]
POV-Ray statistics for finished frames:
demo.pov Statistics (Partial Image Rendered), Resolution 1024 x 768
---------------------------------------------------------------------------Pixels:
385024
Samples:
385024
Smpls/Pxl: 1.00
Rays:
660462
Saved:
1298
Max Level: 0/5
---------------------------------------------------------------------------Ray->Shape Intersection
Tests
Succeeded Percentage
---------------------------------------------------------------------------Box
9556641
6395308
66.92
Cone/Cylinder
12742188
2552058
20.03
CSG Intersection
46721356
2094686
4.48
CSG Union
16989584
790196
4.65
Sphere
21236980
3092195
14.56
True Type Font
25484376
690290
2.71
---------------------------------------------------------------------------Calls to Noise:
0
Calls to DNoise:
173563
---------------------------------------------------------------------------Shadow Ray Tests:
5234905
Succeeded:
44974
Reflected Rays:
275438
---------------------------------------------------------------------------Smallest Alloc:
32 bytes
Largest:
20508
Peak memory used:
925085 bytes
---------------------------------------------------------------------------Time For Trace:
0 hours 15 minutes
6.0 seconds (906 seconds)
Total Time:
0 hours 15 minutes
6.0 seconds (906 seconds)
[beowulf@node8 povray31]$ exit
Script done on Sat Nov 24 03:02:49 2001
Poolballs on 3 nodes
Script started on Sat Nov 24 02:21:11 2001
[beowulf@node8 povray31]$ mpirun n0 n0-2 ./mpi-x-povray -i demo.pov -w1024 -h768 _ Persistence
of Vision(tm) Ray Tracer Version 3.1g
This is an unofficial version compiled by:
Leon Verrall ([email protected])
The POV-Ray Team(tm) is not responsible for supporting this version.
Copyright 1999 POV-Ray Team(tm)
Initializing MPI-POVRAY
Parsing (Slave PE 1).
Parsing (Slave PE 2)......
Parsing (Slave PE 3)......
demo.pov:26: warning: All #version and #declares of float, vector, and color require semicolon ';' at end.
: All #version and #declares of float, vector, and color require semi-colon ';' at end.
13_Pball.inc:98: warning: All #version and #declares of float, vector, and color require semicolon ';' at end.
No pigment type given.
Creating bounding slabs.No pigment type given.
Scene contains 13 frame level objects; 0 infinite.
Rendering...(Slave PE 3)
Creating bounding slabs..
Scene contains 13 frame level objects; 0 infinite.
Rendering...(Slave PE 2)
Creating bounding slabs.
Scene contains 13 frame level objects; 0 infinite.
Rendering...(Slave PE 1)
.....All blocks are assigned. Stopping PE 2.
Done Tracing
Slave PE 2 has exited. 2 minions left...
.........All blocks are assigned. Stopping PE 3.
Done Tracing
Slave PE 3 has exited. 1 minions left...
Cluster Computing
Page 141
Revision Version 2.3
Copyright © 2003
.....................All blocks are assigned. Stopping PE 1.
Done Tracing
Slave PE 1 has exited. 0 minions left...
All slave tasks have exited!
PE Distribution Statistics:
Slave PE [ done ]
1 [31.38%]
2 [34.24%]
Slave PE
[ done ]
3
[34.38%]
POV-Ray statistics for finished frames:
demo.pov Statistics (Partial Image Rendered), Resolution 1024 x 768
---------------------------------------------------------------------------Pixels:
246784
Samples:
246784
Smpls/Pxl: 1.00
Rays:
430194
Saved:
2213
Max Level: 0/5
---------------------------------------------------------------------------Ray->Shape Intersection
Tests
Succeeded Percentage
---------------------------------------------------------------------------Box
6235065
4159613
66.71
Cone/Cylinder
8313420
1677537
20.18
CSG Intersection
30482540
1396883
4.58
CSG Union
11084560
532510
4.80
Sphere
13855700
2036300
14.70
True Type Font
16626840
478098
2.88
---------------------------------------------------------------------------Calls to Noise:
0
Calls to DNoise:
113900
---------------------------------------------------------------------------Shadow Ray Tests:
3442452
Succeeded:
32137
Reflected Rays:
183410
---------------------------------------------------------------------------Smallest Alloc:
32 bytes
Largest:
20508
Peak memory used:
1089009 bytes
---------------------------------------------------------------------------Time For Trace:
0 hours 10 minutes
5.0 seconds (605 seconds)
Total Time:
0 hours 10 minutes
5.0 seconds (605 seconds)
[beowulf@node8 povray31]$ exit
Script done on Sat Nov 24 02:33:58 2001
Poolballs on 4 nodes
Script started on Sat Nov 24 02:12:05 2001
[beowulf@node8 povray31]$ mpirun n0 n0-3 ./mpi-x-povray -i demo.pov -w1024 -h768 _ Persistence
of Vision(tm) Ray Tracer Version 3.1g
This is an unofficial version compiled by:
Leon Verrall ([email protected])
The POV-Ray Team(tm) is not responsible for supporting this version.
Copyright 1999 POV-Ray Team(tm)
Initializing MPI-POVRAY
Parsing (Slave PE 1)..
Parsing (Slave PE 2)
Parsing (Slave PE 3).......
Parsing (Slave PE 4).........
demo.pov:26: warning: All #version and #declares of float, vector, and color require semicolon ';' at end.
13_Pball.inc:100: warning: All #version and #declares of float, vector, and color require
semi-colon ';' at end.
No pigment type given.
Rendering...(Slave PE 1)
Creating bounding slabs.
Scene contains 13 frame level objects; 0 infinite.
No pigment type given.
Creating bounding slabs.
Rendering...(Slave PE 2)
..
Scene contains 13 frame level objects; 0 infinite.
Rendering...(Slave PE 4)
Page 142
Cluster Computing
Revision Version 2.3
Copyright © 2003
......13_Pball.inc:100: warning: All #version and #declares of float, vector, and color
require semi-colon ';' at end.
13_Pball.inc:101: warning: All #version and #declares of float, vector, and color require
semi-colon ';' at end.
No pigment type given.
......No pigment type given.
No pigment type given.
No pigment type given.
......No pigment type given.
Creating bounding slabs.......
Scene contains 13 frame level objects; 0 infinite.
Rendering...(Slave PE 3)
....All blocks are assigned. Stopping PE 3.
Done Tracing
Slave PE 3 has exited. 3 minions left...
....All blocks are assigned. Stopping PE 4.
Done Tracing
Slave PE 4 has exited. 2 minions left...
...........All blocks are assigned. Stopping PE 2.
Done Tracing
Slave PE 2 has exited. 1 minions left...
..................All blocks are assigned. Stopping PE 1.
Done Tracing
Slave PE 1 has exited. 0 minions left...
All slave tasks have exited!
PE Distribution Statistics:
Slave PE [ done ]
1 [22.92%]
2 [25.78%]
4 [25.78%]
Slave PE
[ done ]
3
[25.52%]
POV-Ray statistics for finished frames:
demo.pov Statistics (Partial Image Rendered), Resolution 1024 x 768
---------------------------------------------------------------------------Pixels:
180224
Samples:
180224
Smpls/Pxl: 1.00
Rays:
315611
Saved:
966
Max Level: 0/5
---------------------------------------------------------------------------Ray->Shape Intersection
Tests
Succeeded Percentage
---------------------------------------------------------------------------Box
4586931
3050974
66.51
Cone/Cylinder
6115908
1210434
19.79
CSG Intersection
22424996
1030496
4.60
CSG Union
8154544
391033
4.80
Sphere
10193180
1511995
14.83
True Type Font
12231816
339630
2.78
---------------------------------------------------------------------------Calls to Noise:
0
Calls to DNoise:
82219
---------------------------------------------------------------------------Shadow Ray Tests:
2535182
Succeeded:
23202
Reflected Rays:
135387
---------------------------------------------------------------------------Smallest Alloc:
32 bytes
Largest:
20508
Peak memory used:
1089029 bytes
---------------------------------------------------------------------------Time For Trace:
0 hours 7 minutes 37.0 seconds (457 seconds)
Total Time:
0 hours 7 minutes 37.0 seconds (457 seconds)
[beowulf@node8 povray31]$ exit
Script done on Sat Nov 24 02:21:03 2001
Cluster Computing
Page 143
Revision Version 2.3
Copyright © 2003
Poolballs on 5 nodes
Script started on Sat Nov 24 02:04:32 2001
[beowulf@node8 povray31]$ mpirui__[Kn n0 n0-4 ./mpi-x-run__[K__[K__[Kpovray -i demo.pov -w1024
-h768 _ Persistence of Vision(tm) Ray Tracer Version 3.1g
This is an unofficial version compiled by:
Leon Verrall ([email protected])
The POV-Ray Team(tm) is not responsible for supporting this version.
Copyright 1999 POV-Ray Team(tm)
Initializing MPI-POVRAY
Parsing (Slave PE 1).
Parsing (Slave PE 4)..
Parsing (Slave PE 3)..
Parsing (Slave PE 2)........
Parsing (Slave PE 5).......
demo.pov:26: warning: All #version and #declares of float, vector, and color require semicolon ';' at end.
13_Pball.inc:98: warning: All #version and #declares of float, vector, and color require semicolon ';' at end.
No pigment type given.
Rendering...(Slave PE 2)
Creating bounding slabs..No pigment type given.
Scene contains 13 frame level objects; 0 infinite.
No pigment type given.
.13_Pball.inc:100: warning: All #version and #declares of float, vector, and color require
semi-colon ';' at end.
No pigment type given.
.
Rendering...(Slave PE 4)
Creating bounding slabs..
Scene contains 13 frame level objects; 0 infinite.
.No pigment type given.
Rendering...(Slave PE 5)
Creating bounding slabs.........
Scene contains 13 frame level objects; 0 infinite.
Rendering...(Slave PE 3)
.........................No pigment type given.
Creating bounding slabs.
Scene contains 13 frame level objects; 0 infinite.
Rendering...(Slave PE 1)
...All blocks are assigned. Stopping PE 4.
Done Tracing
Slave PE 4 has exited. 4 minions left...
....................All blocks are assigned. Stopping PE 5.
Done Tracing
Slave PE 5 has exited. 3 minions left...
.......................All blocks are assigned. Stopping PE 1.
Done Tracing
Slave PE 1 has exited. 2 minions left...
..All blocks are assigned. Stopping PE 3.
.
Done Tracing
Slave PE 3 has exited. 1 minions left...
...............All blocks are assigned. Stopping PE 2.
Done Tracing
Slave PE 2 has exited. 0 minions left...
All slave tasks have exited!
PE Distribution Statistics:
Slave PE [ done ]
1 [17.71%]
Page 144
Slave PE
[ done ]
Cluster Computing
Revision Version 2.3
Copyright © 2003
2
4
[20.44%]
[20.44%]
3
5
[20.70%]
[20.70%]
POV-Ray statistics for finished frames:
demo.pov Statistics (Partial Image Rendered), Resolution 1024 x 768
---------------------------------------------------------------------------Pixels:
160768
Samples:
160768
Smpls/Pxl: 1.00
Rays:
274492
Saved:
707
Max Level: 0/5
---------------------------------------------------------------------------Ray->Shape Intersection
Tests
Succeeded Percentage
---------------------------------------------------------------------------Box
3977136
2654137
66.73
Cone/Cylinder
5302848
1078482
20.34
CSG Intersection
19443776
884746
4.55
CSG Union
7070464
339164
4.80
Sphere
8838080
1267375
14.34
True Type Font
10605696
305770
2.88
---------------------------------------------------------------------------Calls to Noise:
0
Calls to DNoise:
73461
---------------------------------------------------------------------------Shadow Ray Tests:
2185547
Succeeded:
18978
Reflected Rays:
113724
---------------------------------------------------------------------------Smallest Alloc:
32 bytes
Largest:
20508
Peak memory used:
1089049 bytes
---------------------------------------------------------------------------Time For Trace:
0 hours 6 minutes
4.0 seconds (364 seconds)
Total Time:
0 hours 6 minutes
4.0 seconds (364 seconds)
[beowulf@node8 povray31]$ exit
Script done on Sat Nov 24 02:11:57 2001
Poolballs on 6 nodes
Script started on Sat Nov 24 01:57:28 2001
[beowulf@node8 povray31]$ mpirun n0 n0-5 ./mpi-x-povray -i demo.pov -w1024 -h 76 _8
_[A_[79C__7688
_[K_[A_[77C768
Persistence of Vision(tm) Ray Tracer Version 3.1g
This is an unofficial version compiled by:
Leon Verrall ([email protected])
The POV-Ray Team(tm) is not responsible for supporting this version.
Copyright 1999 POV-Ray Team(tm)
Initializing MPI-POVRAY
Parsing (Slave PE 1).
Parsing (Slave PE 4)
Parsing (Slave PE 5)....
Parsing (Slave PE 6)....
Parsing (Slave PE 3)......
Parsing (Slave PE 2)............
demo.pov:26: warning: All #version and #declares of float, vector, and color require semicolon ';' at end.
13_Pball.inc:98: warning: All #version and #declares of float, vector, and color require semicolon ';' at end.
No pigment type given.
Creating bounding slabs.
Scene contains 13 frame level objects; 0 infinite.
Rendering...(Slave PE 2)
Creating bounding slabs.
Scene contains 13 frame level objects; 0 infinite.
Creating bounding slabs.
Scene contains 13 frame level objects; 0 infinite.
Rendering...(Slave PE 6)
Rendering...(Slave PE 4)
Creating bounding slabs........
Scene contains 13 frame level objects; 0 infinite.
Rendering...(Slave PE 5)
Cluster Computing
Page 145
Revision Version 2.3
Copyright © 2003
...All blocks are assigned. Stopping PE 6.
Done Tracing
Slave PE 6 has exited. 5 minions left...
.............................All blocks are assigned. Stopping PE 3.
All blocks are assigned. Stopping PE 2.
Done Tracing
Done Tracing
Slave PE 2 has exited. 4 minions left...
.
Slave PE 3 has exited. 3 minions left...
.....................All blocks are assigned. Stopping PE 4.
Done Tracing
Slave PE 4 has exited. 2 minions left...
....................All blocks are assigned. Stopping PE 5.
Done TracingAll blocks are assigned. Stopping PE 1.
Slave PE 5 has exited. 1 minions left...
Done Tracing
Slave PE 1 has exited. 0 minions left...
All slave tasks have exited!
PE Distribution Statistics:
Slave PE [ done ]
1 [12.89%]
2 [16.93%]
4 [17.84%]
6 [17.45%]
Slave PE
[ done ]
3
5
[17.32%]
[17.58%]
POV-Ray statistics for finished frames:
demo.pov Statistics (Partial Image Rendered), Resolution 1024 x 768
---------------------------------------------------------------------------Pixels:
101376
Samples:
101376
Smpls/Pxl: 1.00
Rays:
177122
Saved:
474
Max Level: 0/5
---------------------------------------------------------------------------Ray->Shape Intersection
Tests
Succeeded Percentage
---------------------------------------------------------------------------Box
2611044
1726022
66.10
Cone/Cylinder
3481392
731658
21.02
CSG Intersection
12765104
598613
4.69
CSG Union
4641856
232867
5.02
Sphere
5802320
837760
14.44
True Type Font
6962784
212564
3.05
---------------------------------------------------------------------------Calls to Noise:
0
Calls to DNoise:
50314
---------------------------------------------------------------------------Shadow Ray Tests:
1475084
Succeeded:
12732
Reflected Rays:
75746
---------------------------------------------------------------------------Smallest Alloc:
32 bytes
Largest:
20508
Peak memory used:
1252973 bytes
---------------------------------------------------------------------------Time For Trace:
0 hours 5 minutes
9.0 seconds (309 seconds)
Total Time:
0 hours 5 minutes
9.0 seconds (309 seconds)
[beowulf@node8 povray31]$ exit
Script done on Sat Nov 24 02:03:56 2001
Page 146
Cluster Computing
Revision Version 2.3
Copyright © 2003
Poolballs on 7 nodes
Script started on Sat Nov 24 01:51:13 2001
[beowulf@node8 povray31]$ mpirun n0 n0-6 ./mpi-x-povray -i demo.pov -w1024 -h768 _ Persistence
of Vision(tm) Ray Tracer Version 3.1g
This is an unofficial version compiled by:
Leon Verrall ([email protected])
The POV-Ray Team(tm) is not responsible for supporting this version.
Copyright 1999 POV-Ray Team(tm)
Initializing MPI-POVRAY
Parsing (Slave PE 2)
Parsing (Slave PE 7).
Parsing (Slave PE 1)...
Parsing (Slave PE 4)..
Parsing (Slave PE 3)...
Parsing (Slave PE 6)......
Parsing (Slave PE 5)...............
demo.pov:26: warning: All #version and #declares of float, vector, and color require semicolon ';' at end.
13_Pball.inc:98: warning: All #version and #declares of float, vector, and color require semicolon ';' at end.
No pigment type given.
Creating bounding slabs.
Scene contains 13 frame level objects; 0 infinite.
Rendering...(Slave PE 7)
Rendering...(Slave PE 3)
Rendering...(Slave PE 4)
Rendering...(Slave PE 6)
Rendering...(Slave PE 2)
Rendering...(Slave PE 5)
Rendering...(Slave PE 1)
...All blocks are assigned. Stopping PE 4.
Done Tracing
Slave PE 4 has exited. 6 minions left...
..All blocks are assigned. Stopping PE 6.
..
Done Tracing.
Slave PE 6 has exited. 5 minions left...
..............All blocks are assigned. Stopping PE 3.
All blocks are assigned. Stopping PE 7.
Done Tracing
Done Tracing
Slave PE 3 has exited. 4 minions left...
Slave PE 7 has exited. 3 minions left...
....All blocks are assigned. Stopping PE 5.
Done Tracing..
Slave PE 5 has exited. 2 minions left...
All blocks are assigned. Stopping PE 2.
Done Tracing
Slave PE 2 has exited. 1 minions left...
........All blocks are assigned. Stopping PE 1.
Done Tracing
Slave PE 1 has exited. 0 minions left...
All slave tasks have exited!
PE Distribution Statistics:
Slave PE [ done ]
1 [11.98%]
2 [14.71%]
4 [14.84%]
6 [14.84%]
Slave PE
[ done ]
3
5
7
[14.45%]
[14.58%]
[14.58%]
POV-Ray statistics for finished frames:
demo.pov Statistics (Partial Image Rendered), Resolution 1024 x 768
Cluster Computing
Page 147
Revision Version 2.3
Copyright © 2003
---------------------------------------------------------------------------Pixels:
94208
Samples:
94208
Smpls/Pxl: 1.00
Rays:
164181
Saved:
418
Max Level: 0/5
---------------------------------------------------------------------------Ray->Shape Intersection
Tests
Succeeded Percentage
---------------------------------------------------------------------------Box
2395890
1608816
67.15
Cone/Cylinder
3194520
656058
20.54
CSG Intersection
11713240
558802
4.77
CSG Union
4259360
217846
5.11
Sphere
5324200
810185
15.22
True Type Font
6389040
207192
3.24
---------------------------------------------------------------------------Calls to Noise:
0
Calls to DNoise:
44546
---------------------------------------------------------------------------Shadow Ray Tests:
1331811
Succeeded:
11470
Reflected Rays:
69973
---------------------------------------------------------------------------Smallest Alloc:
32 bytes
Largest:
20508
Peak memory used:
1089089 bytes
---------------------------------------------------------------------------Time For Trace:
0 hours 4 minutes 22.0 seconds (262 seconds)
Total Time:
0 hours 4 minutes 22.0 seconds (262 seconds)
[beowulf@node8 povray31]$ exit
Script done on Sat Nov 24 01:56:59 2001
Poolballs on 8 nodes
Script started on Sun Nov 25 18:19:13 2001
[beowulf@node8 povray31]$ mpirun ./mpi-x-povray -i demo.pov -h1024 h768____________[P1__[1@w1_1024 -h768 mpirun: cannot parse: Exec format error
[beowulf@node8 povray31]$ mpirun ./mpi-x-povray -i demo.pov -w1024 h768________________________________________[1@n.__[1@0.__[1@ .__[1@n.__[1@0.__[1@.__[1@7.__[1@ .__[38C8 Persistence of Vision(tm) Ray Tracer Version 3.1g
This is an unofficial version compiled by:
Leon Verrall ([email protected])
The POV-Ray Team(tm) is not responsible for supporting this version.
Copyright 1999 POV-Ray Team(tm)
Initializing MPI-POVRAY
Parsing
Parsing
Parsing
Parsing
Parsing
Parsing
Parsing
Parsing
(Slave
(Slave
(Slave
(Slave
(Slave
(Slave
(Slave
(Slave
PE
PE
PE
PE
PE
PE
PE
PE
1)
4)
2)
3)
6)
8)
7)
5).....................
13_Pball.inc:101: warning: All #version and #declares of float, vector, and color require
semi-colon ';' at end.
Creating bounding slabs.No pigment type given.
No pigment type given.
No pigment type given.
Scene contains 13 frame level objects; 0 infinite.
Rendering...(Slave PE 8)
Rendering...(Slave PE 3)
Rendering...(Slave PE 7)
Rendering...(Slave PE 1)
Rendering...(Slave PE 2)
Rendering...(Slave PE 4)
Rendering...(Slave PE 6)
Rendering...(Slave PE 5)
..All blocks are assigned. Stopping PE 7.
Done Tracing
Slave PE 7 has exited. 7 minions left...
...All blocks are assigned. Stopping PE 5.
Page 148
Cluster Computing
Revision Version 2.3
Copyright © 2003
Done Tracing
Slave PE 5 has exited. 6 minions left...
.........................All blocks are assigned. Stopping PE 2.
Done Tracing..
Slave PE 2 has exited. 5 minions left...
All blocks are assigned. Stopping PE 6.
.
Done Tracing
Slave PE 6 has exited. 4 minions left...
..............................All blocks are assigned. Stopping PE 4.
Done Tracing
Slave PE 4 has exited. 3 minions left...
........All blocks are assigned. Stopping PE 3.
Done Tracing
Slave PE 3 has exited. 2 minions left...
....All blocks are assigned. Stopping PE 8.
Done Tracing
Slave PE 8 has exited. 1 minions left...
..............All blocks are assigned. Stopping PE 1.
Done Tracing
Slave PE 1 has exited. 0 minions left...
All slave tasks have exited!
PE Distribution Statistics:
Slave PE [ done ]
1 [10.55%]
2 [12.89%]
4 [12.63%]
6 [12.37%]
8 [13.15%]
Slave PE
[ done ]
3
5
7
[12.89%]
[12.63%]
[12.89%]
POV-Ray statistics for finished frames:
demo.pov Statistics (Partial Image Rendered), Resolution 1024 x 768
---------------------------------------------------------------------------Pixels:
82944
Samples:
82944
Smpls/Pxl: 1.00
Rays:
147318
Saved:
257
Max Level: 0/5
---------------------------------------------------------------------------Ray->Shape Intersection
Tests
Succeeded Percentage
---------------------------------------------------------------------------Box
2134350
1424997
66.76
Cone/Cylinder
2845800
559332
19.65
CSG Intersection
10434600
481141
4.61
CSG Union
3794400
183495
4.84
Sphere
4743000
711005
14.99
True Type Font
5691600
164144
2.88
---------------------------------------------------------------------------Calls to Noise:
0
Calls to DNoise:
37254
---------------------------------------------------------------------------Shadow Ray Tests:
1171157
Succeeded:
9252
Reflected Rays:
64374
---------------------------------------------------------------------------Smallest Alloc:
32 bytes
Largest:
20508
Peak memory used:
1089109 bytes
---------------------------------------------------------------------------Time For Trace:
0 hours 3 minutes 49.0 seconds (229 seconds)
Total Time:
0 hours 3 minutes 49.0 seconds (229 seconds)
[beowulf@node8 povray31]$ exit
Script done on Sun Nov 25 18:26:36 2001
Cluster Computing
Page 149
Revision Version 2.3
Copyright © 2003
Tulips on 1 node
Script started on Wed Nov 21 01:48:42 2001
[beowulf@node1 povray31]$ x-povray -i tulips.pov Persistence of Vision(tm) Ray Tracer Version
3.1g.Linux.gcc
This is an official version prepared by the POV-Ray Team(tm). See the
documentation on how to contact the authors or visit us on the
internet at http://www.povray.org.
Copyright 1999 POV-Ray Team(tm)
Parsing Options
Input file: tulips.pov (compatible to version 3.1)
Remove bounds........On Split unions........Off
Library paths: /usr/local/lib/povray31 /usr/local/lib/povray31/include
Output Options
Image resolution 320 by 240 (rows 1 to 240, columns 1 to 320).
Output file: tulips.png, 24 bpp PNG
Graphic display.....Off
Mosaic preview......Off
CPU usage histogram.Off
Continued trace.....Off Allow interruption...On Pause when done.....Off
Verbose messages....Off
Tracing Options
Quality: 9
Bounding boxes.......On Bounding threshold: 25
Light Buffer.........On Vista Buffer.........On
Antialiasing........Off
Radiosity...........Off
Animation Options
Clock value....
0.000 (Animation off)
Redirecting Options
All Streams to console..........On
Debug Stream to console.........On
Fatal Stream to console.........On
Render Stream to console........On
Statistics Stream to console....On
Warning Stream to console.......On
Parsing.......
Focal blur is used. Standard antialiasing is switched off.
Creating bounding slabs.
Scene contains 22011 frame level objects; 1 infinite.
Creating light buffers.......
Rendering... Done Tracing
tulips.pov Statistics, Resolution 320 x 240
---------------------------------------------------------------------------Pixels:
76800
Samples:
19660800
Smpls/Pxl: 256.00
Rays:
19660800
Saved:
0
Max Level: 1/5
---------------------------------------------------------------------------Ray->Shape Intersection
Tests
Succeeded Percentage
---------------------------------------------------------------------------Blob
1275777783
298321419
23.38
Blob Component
2381289467
2259449386
94.88
Blob Bound
2551555566
2381289467
93.33
Box
10251883237
2972673609
29.00
Cone/Cylinder
53735942
2517948
4.69
CSG Intersection
8516484938
285092797
3.35
Plane
8445288949
3616358569
42.82
Torus
2432060743
789401935
32.46
Torus Bound
2432060743
899809497
37.00
Bounding Object
8909077334
1863108116
20.91
Bounding Box
58535570831
11813741448
20.18
Light Buffer
206031501066
87812364509
42.62
---------------------------------------------------------------------------Roots tested:
1684034445
eliminated:
448286134
Calls to Noise:
25936097
Calls to DNoise:
94313101
---------------------------------------------------------------------------Shadow Ray Tests:
1288804564
Succeeded:
195215845
---------------------------------------------------------------------------Smallest Alloc:
11 bytes
Largest:
176088
Peak memory used:
125446732 bytes
Page 150
Cluster Computing
Revision Version 2.3
Copyright © 2003
---------------------------------------------------------------------------Time For Parse:
0 hours 1 minutes
5.0 seconds (65 seconds)
Time For Trace:
77 hours 33 minutes 19.0 seconds (279199 seconds)
Total Time:
77 hours 34 minutes 24.0 seconds (279264 seconds)
[beowulf@node1 povray31]$ exit
Script done on Sun Nov 25 17:40:58 2001
Cluster Computing
Page 151
10.
End of Document